This week’s parliamentary elections are widely seen as a referendum on both President Yoon Suk Yeol and his archrival, Lee Jae-myung, the opposition leader.

‘Gladiator Politics’ Dominate Election Season in Polarized South Korea

First minister calls for end to vexatious reports after far-right agitators attempt to ‘overwhelm’ official systems

Police spammed with complaints by neo-Nazis under new Scottish hate crime law | Hate crime | The Guardian

Hovering over a state visit to Washington is the possibility of a swing in American foreign policy if Donald Trump returns to the White House.

Japan and US Seek to Strengthen Ties as Kishida Visits - The New York Times

DURHAM, N.H. – Sophomores Macy Shultz (Adams, N.Y.) and Grace Lydon (Rowley, Mass.) each scored a career-high five points, as the UMass Lowell women&#x27;s lacrosse team (10-3, 1-2 AE) took down the New Hampshire Wildcats (2-9, 0-3 AE), 18-10, on Saturday afternoon at Tucker Field.

River Hawks Take Down New Hampshire, 18-10, For First Conference Win - UMass Lowell Athletics

Brian Dorsey, 52, is set to be executed on Tuesday for murdering his cousin and her husband in 2006.

Missouri death row inmate facing 'surgery without anesthesia': lawyer

PORTLAND, Maine (AP) — Power restoration efforts in Maine and New Hampshire entered another day on Saturday with more than 100,000 homes and businesses still in the dark several days after a storm brought powerful winds and heavy, wet snow.

Tens of Thousands Still Without Power Following Powerful Nor'easter in New England

Meteorologists now feel increasingly confident in models that show some parts of the country looking golden for the eclipse on Monday, April 8. Other parts of the country, though, may end up in a cloudy bust.

Eclipse cloud cover forecasts and maps show where skies will clear up for April 8's celestial show - CBS News

Actor Cole Brings Plenty, who had been missing since Sunday amid a domestic violence investigation, has been found dead, authorities in Kansas said Friday.

Missing '1923' actor Cole Brings Plenty, 27, found dead in Kansas

Relatives say the explosion was caused by a tree falling onto a propane tank.

Grandmother killed, girl injured in New Hampshire house explosion during snowstorm - CBS Boston

A nor’easter that dumped more than a foot of snow across the region, causing blackouts and flight cancellations, is expected to weaken on Friday, the National Weather Service said.

Thousands in Maine, New Hampshire and Vermont Without Power After Snowstorm - The New York Times

If an individual is sleep deprived, at what minimum can certain functions continue without getting any sleep? If someone sees something from a distance, but it is just clear enough to be roughly interpreted by the mind in terms of kind, but not specificity, is there an input configuration requirement to guarantee an optimum interpretation of the external world by the human mind? If functions are continuously operating at minimums, could there be a likelihood that some of the factors responsible for optimum outputs may start getting the message that they are no longer necessary? If this is the case, could it cascade into a serious condition? Neurons are known to have around 50 billion proteins. Protein misfolding has been established as a major cause of some neurodegenerative diseases. A possible reason for protein misfolding has been identified as [cellular] environmental stress. A theoretical basis for why environmental stress may be responsible could be to look at working minimums and maximums, for functions, over certain intervals. Theorizing protein folding requires a top-down approach, which means that there is a source for which problems might arise, resulting--eventually--in misfolding and degenerative diseases. The closest destination for the top option is the electrical and chemical signals of neurons. It is theorized that the human mind is responsible for all functions that the brain is said to do. The human mind is the collection of all the electrical and chemical signals, with their interactions and features, in sets, in clusters of neurons, across the central and peripheral nervous system. This means that the human mind is the signals with their interactions and features. All the configurations for which representations and interpretations are based, for interoception and exteroception, have just one possible option in the brain, the signals, not genes, proteins, neurons, or other types of cells. This means that whatever is interpreted, accurately or not, or seems represented, accurately or not, is a result of the signals, such that with signals, interpretations, and features, in sets, are at the top, though their work has implications for other cells in the body, and several other molecules, especially proteins, abundant in neurons. Why? Exercise, for example, is an activity where electrical and chemical signals are interacting actively, with varying features applying. This top activity results in relays of electrical signals that, conceptually, for interactions with chemical signals, in sets, have an effect on other parts of the body. Simply, electrical and chemical signals are involved in every function. The consequences of those functions can also be seen elsewhere. Proteins, Neurons, and Signals Since neurons are not the human mind, or their proteins, because they do not organize information, it can be theorized that as neurons are involved in clusters, with sets of signals, their proteins assist several interactive processes between electrical and chemical signals. This indicates that when there are extended minimums or maximums, in sets of signals, there are ways that they come to have an effect on some of the proteins of the neurons involved in respective clusters [hence environmental stress]. Simply, neurons have proteins, and these proteins contribute to the primary role of signals, but as signals carry out their roles, depending on the pace or direction, they come to also have effects on the protein of some neurons, which may work more or work less, over an interval, which may, for some cases lead to misfolding, and then accumulations and disease, conceptually. It means that the mechanisms of signals can have an effect, conceptually, on the misfolding of proteins, then possibly developing into a degenerative disease for some. Tracking Possible Paths for Protein Misfolding From Signals It is possible to develop a number of minimums, maximums, and interactive-feature stretches of signals, with an exploration of how, for some people with Amyloidosis, that may have resulted from cellular [environmental] stress, to look at the roles that signals played. The interaction of signals is the way electrical signals strike at chemical signals, while the features are some of the states of the signals, in sets, at the times of the interactions. If a person drinks cold water but does not feel it as cold, why did the signals not give the experience of cold? There are several examples like this, but what can be looked at is the intensity of certain experiences or the longevity of others, which may have caused the larger protein folding error. Theorizing why proteins misfold may have to center around the major functional operators the electrical and chemical signals, which can then be sought, as the source toward reversal or prevention. There is a recent story in Nature, AI protein-prediction tool AlphaFold3 is now open source, stating that, "AlphaFold3, unlike its predecessors, is capable of modelling proteins in concert with other molecules. But instead of releasing its underlying code — as was done with AlphaFold2 — DeepMind provided access via a web server that restricted the number and types of predictions scientists could make. Crucially, the AlphaFold3 server prevented scientists from predicting how proteins behave in the presence of potential drugs. But now, DeepMind’s decision to release the code means academic scientists can predict such interactions by running the model themselves. The company initially said that making AlphaFold3 available only through a web server struck the right balance between enabling access for research and protecting commercial ambitions. Isomorphic Labs, a DeepMind spinoff company in London, is applying AlphaFold3 to drug discovery." There is a recent press release, Argonne team breaks new ground in AI-driven protein design, stating that, "Harnessing the power of artificial intelligence (AI) and the world’s fastest supercomputers, a research team led by the U.S. Department of Energy’s (DOE) Argonne National Laboratory has developed an innovative computing framework to speed up the design of new proteins. On the heels of this year’s Nobel Prize in Chemistry, which recognized advances in computational protein design, Argonne’s AI-driven approach has been selected as a finalist for the prestigious Gordon Bell Prize. Presented by the Association of Computing Machinery, the annual prize recognizes breakthroughs in using high performance computing to solve complex science problems. Mapping a protein’s amino acid sequence to its structure and function is a long-standing research challenge. Each unique arrangement of amino acids — the building blocks of proteins — can yield different properties and behaviors. The sheer volume of potential variations makes it impractical to test them all through experiments alone. To put this in perspective, modifying just three amino acids in a sequence of 20 creates 8,000 possible combinations. But most proteins are far more complex, with some research targets containing hundreds to thousands of amino acids." If an individual is sleep deprived, at what minimum can certain functions continue without getting any sleep? If someone sees something from a distance, but it is just clear enough to be roughly interpreted by the mind in terms of kind, but not specificity, is there an input configuration requirement to guarantee an optimum interpretation of the external world by the human mind? If functions are continuously operating at minimums, could there be a likelihood that some of the factors responsible for optimum outputs may start getting the message that they are no longer necessary? If this is the case, could it cascade into a serious condition? Neurons are known to have around 50 billion proteins . Protein misfolding has been established as a major cause of some neurodegenerative diseases . A possible reason for protein misfolding has been identified as [cellular] environmental stress . A theoretical basis for why environmental stress may be responsible could be to look at working minimums and maximums, for functions, over certain intervals. have around 50 billion proteins major cause of some neurodegenerative diseases as [cellular] environmental stress Theorizing protein folding requires a top-down approach, which means that there is a source for which problems might arise, resulting--eventually--in misfolding and degenerative diseases. The closest destination for the top option is the electrical and chemical signals of neurons. It is theorized that the human mind is responsible for all functions that the brain is said to do. The human mind is the collection of all the electrical and chemical signals, with their interactions and features, in sets, in clusters of neurons, across the central and peripheral nervous system. This means that the human mind is the signals with their interactions and features. All the configurations for which representations and interpretations are based, for interoception and exteroception, have just one possible option in the brain, the signals, not genes, proteins, neurons, or other types of cells. This means that whatever is interpreted, accurately or not, or seems represented, accurately or not, is a result of the signals, such that with signals, interpretations, and features, in sets, are at the top, though their work has implications for other cells in the body, and several other molecules, especially proteins, abundant in neurons. Why? Exercise, for example, is an activity where electrical and chemical signals are interacting actively, with varying features applying. This top activity results in relays of electrical signals that, conceptually, for interactions with chemical signals, in sets, have an effect on other parts of the body. Simply, electrical and chemical signals are involved in every function. The consequences of those functions can also be seen elsewhere. Proteins, Neurons, and Signals Since neurons are not the human mind, or their proteins, because they do not organize information, it can be theorized that as neurons are involved in clusters, with sets of signals, their proteins assist several interactive processes between electrical and chemical signals. This indicates that when there are extended minimums or maximums, in sets of signals, there are ways that they come to have an effect on some of the proteins of the neurons involved in respective clusters [hence environmental stress]. Simply, neurons have proteins, and these proteins contribute to the primary role of signals, but as signals carry out their roles, depending on the pace or direction, they come to also have effects on the protein of some neurons, which may work more or work less, over an interval, which may, for some cases lead to misfolding, and then accumulations and disease, conceptually. It means that the mechanisms of signals can have an effect, conceptually, on the misfolding of proteins, then possibly developing into a degenerative disease for some. Tracking Possible Paths for Protein Misfolding From Signals It is possible to develop a number of minimums, maximums, and interactive-feature stretches of signals, with an exploration of how, for some people with Amyloidosis, that may have resulted from cellular [environmental] stress, to look at the roles that signals played. The interaction of signals is the way electrical signals strike at chemical signals , while the features are some of the states of the signals, in sets, at the times of the interactions. If a person drinks cold water but does not feel it as cold, why did the signals not give the experience of cold? There are several examples like this, but what can be looked at is the intensity of certain experiences or the longevity of others, which may have caused the larger protein folding error. The interaction of signals is the way electrical signals strike at chemical signals Theorizing why proteins misfold may have to center around the major functional operators the electrical and chemical signals, which can then be sought, as the source toward reversal or prevention. There is a recent story in Nature , AI protein-prediction tool AlphaFold3 is now open source , stating that, "AlphaFold3, unlike its predecessors, is capable of modelling proteins in concert with other molecules. But instead of releasing its underlying code — as was done with AlphaFold2 — DeepMind provided access via a web server that restricted the number and types of predictions scientists could make. Crucially, the AlphaFold3 server prevented scientists from predicting how proteins behave in the presence of potential drugs. But now, DeepMind’s decision to release the code means academic scientists can predict such interactions by running the model themselves. The company initially said that making AlphaFold3 available only through a web server struck the right balance between enabling access for research and protecting commercial ambitions. Isomorphic Labs, a DeepMind spinoff company in London, is applying AlphaFold3 to drug discovery." Nature AI protein-prediction tool AlphaFold3 is now open source There is a recent press release, Argonne team breaks new ground in AI-driven protein design , stating that, "Harnessing the power of artificial intelligence (AI) and the world’s fastest supercomputers, a research team led by the U.S. Department of Energy’s (DOE) Argonne National Laboratory has developed an innovative computing framework to speed up the design of new proteins. On the heels of this year’s Nobel Prize in Chemistry, which recognized advances in computational protein design, Argonne’s AI-driven approach has been selected as a finalist for the prestigious Gordon Bell Prize. Presented by the Association of Computing Machinery, the annual prize recognizes breakthroughs in using high performance computing to solve complex science problems. Mapping a protein’s amino acid sequence to its structure and function is a long-standing research challenge. Each unique arrangement of amino acids — the building blocks of proteins — can yield different properties and behaviors. The sheer volume of potential variations makes it impractical to test them all through experiments alone. To put this in perspective, modifying just three amino acids in a sequence of 20 creates 8,000 possible combinations. But most proteins are far more complex, with some research targets containing hundreds to thousands of amino acids." Argonne team breaks new ground in AI-driven protein design

AlphaFold 3: How Does Proteins Misfolding Cause Neurodegenerative Diseases?

Predicting Links in Graphs with Graph Neural Networks and DGL.ai

አገናኞችን በግራፍ መተንበይ ከግራፍ ነርቭ ኔትወርኮች እና DGL.ai ጋር

Predikce odkazů v grafech pomocí Graph Neuron Networks a DGL.ai

Forudsigelse af links i grafer med Graph Neural Networks og DGL.ai

Predicción de vínculos en gráficos con redes neuronales gráficas y DGL.ai

Grafikoetan loturak aurreikustea Graph Neural Networks eta DGL.ai

グラフニューラルネットワークと DGL.ai によるグラフ内のリンクの予測

ទស្សន៍ទាយតំណភ្ជាប់ក្នុងក្រាហ្វជាមួយបណ្តាញសរសៃប្រសាទក្រាហ្វ និង DGL.ai

Графиктик нейрон тармактары жана DGL.ai менен графиктердеги шилтемелерди болжолдоо

Saišu prognozēšana grafikos, izmantojot grafiku neironu tīklus un DGL.ai

Maminavina rohy amin'ny sary miaraka amin'ny Graph Neural Networks sy DGL.ai

Пешгӯии пайвандҳо дар графикҳо бо шабакаҳои графикии нейронӣ ва DGL.ai

การคาดการณ์ลิงก์ในกราฟด้วย Graph Neural Networks และ DGL.ai

Authors:
(1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk;
(2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction 2 Related Work 3 Framework 4 Main Results 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 11 Other Potential Applications 8 Proof of Theorem 2 Before diving into the detailed proof, we present an outline of the structure of the proof and prove a lemma which we use in the proof of the theorem. Outline of the proof for the thereom 1. Suppose we are given Vi and Vj , two test groups which satisfy the premise of the theorem; 2. Then, we can approximate and bound the loss of each vertex in these groups based on the nearest vertex in the training set by extending the result from Theorem 1; 3. If we can show that there exists a constant independent of the property of each test group, then we obtain the results of Theorem 2. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Authors: Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Framework 3 Framework 4 Main Results 4 Main Results 5 A Case Study on Shortest-Path Distance 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 7 Proof of Theorem 1 8 Proof of Theorem 2 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 10 Additional Experiments Details and Results 11 Other Potential Applications 11 Other Potential Applications 8 Proof of Theorem 2 Before diving into the detailed proof, we present an outline of the structure of the proof and prove a lemma which we use in the proof of the theorem. Outline of the proof for the thereom Outline of the proof for the thereom 1. Suppose we are given Vi and Vj , two test groups which satisfy the premise of the theorem; 2. Then, we can approximate and bound the loss of each vertex in these groups based on the nearest vertex in the training set by extending the result from Theorem 1; 3. If we can show that there exists a constant independent of the property of each test group, then we obtain the results of Theorem 2. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Unfair Generalization in Graph Neural Networks (GNNs)

Authors:
(1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk;
(2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction 2 Related Work 3 Framework 4 Main Results 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 11 Other Potential Applications 4 Main Results In this section, we present the main results derived from our proposed framework for studying the generalization and fairness of GNNs with respect to different structural groups in a graph. We start by demonstrating the relationship between the generalization performance of the model and the structural distance between the subgroups and the training set (Theorem 1). We then show that such a structural relation can induce disparities in the generalization performance across different subgroups (Theorem 2). Finally, we connect these theoretical results to the recent deep learning spline line theory [2] to provide additional insights. Detailed proof of the results presented in this paper can be found in the supplementary material. 4.1 Generalization Performance and Structural Distance Theorem 1 states that the generalization risk of a classifier with respect to a structural subgroup depends on the empirical training loss of the classifier, the structural distance between the subgroup and the training set, and the distortion (topology awareness) of GNNs. Such a relation provides a new perspective on understanding the link between topology awareness and generalization performance in GNNs. Prior research in this area has mainly focused on the connection between GNNs and the Weisfeiler-Lehman algorithm, emphasizing the ability of GNNs to differentiate various graph structures. However, Theorem 1 highlights a direct link between topology awareness and generalization performance, and a potential issue of fairness in GNNs, as it shows that the model’s performance on different structural subgroups is influenced by their distance to the training set. Next, we extend the result of Theorem 1 to analyze the relation between the generalization of different structural subgroups and their distance to the training set. 4.2 Unfair Generalization Performance on Structural Subgroups Theorem 2 states that when comparing two subgroups, if the difference in their structural distance to the training set is larger than the threshold (δα), then GNNs exhibit unfair generalization performance with respect to these groups. Specifically, the theorem states that GNNs generalize better on subgroups that are structurally closer to the training set, which leads to accuracy disparities among different subgroups. Additionally, the theorem notes that the topology awareness of GNNs (measured by the distortion rate α) plays a key role in determining the strength of this accuracy disparity. This has important implications for the use of GNNs in critical applications, as it suggests that increased topology awareness in GNNs can lead to improved generalization performance but also a greater potential for unfairness in terms of accuracy disparities among different structural subgroups. As such, it is important to carefully consider the trade-offs between topology awareness and fairness when designing new GNNs and applying them to real-world problems. 4.3 Connection with Deep Learning Spline Theory Our theoretical findings reveal that when considering the performance of the model from the viewpoint of the embedding space, the training set creates a “valley” effect on generalization performance, with the training vertex at the bottom of the valley. Representations that are closer to the embedding of the training set perform better, while those farther away perform worse. This pattern has an interesting connection with deep learning spline theory [2]. This theory suggests that in DNNs, the learning process for classification divides the input/latent space into small circles (referred to as power diagrams), where representations within each circle are grouped, and representations closer to the centre of the circle are predicted with more confidence. Our findings suggest that in GNNs, the mapping of test data to these power diagrams is related to the structural similarity between test and training data in the graph. This connection provides further validation for our derived results. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Authors: Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Framework 3 Framework 4 Main Results 4 Main Results 5 A Case Study on Shortest-Path Distance 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 7 Proof of Theorem 1 8 Proof of Theorem 2 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 10 Additional Experiments Details and Results 11 Other Potential Applications 11 Other Potential Applications 4 Main Results In this section, we present the main results derived from our proposed framework for studying the generalization and fairness of GNNs with respect to different structural groups in a graph. We start by demonstrating the relationship between the generalization performance of the model and the structural distance between the subgroups and the training set (Theorem 1). We then show that such a structural relation can induce disparities in the generalization performance across different subgroups (Theorem 2). Finally, we connect these theoretical results to the recent deep learning spline line theory [2] to provide additional insights. Detailed proof of the results presented in this paper can be found in the supplementary material. 4.1 Generalization Performance and Structural Distance Theorem 1 states that the generalization risk of a classifier with respect to a structural subgroup depends on the empirical training loss of the classifier, the structural distance between the subgroup and the training set, and the distortion (topology awareness) of GNNs. Such a relation provides a new perspective on understanding the link between topology awareness and generalization performance in GNNs. Prior research in this area has mainly focused on the connection between GNNs and the Weisfeiler-Lehman algorithm, emphasizing the ability of GNNs to differentiate various graph structures. However, Theorem 1 highlights a direct link between topology awareness and generalization performance, and a potential issue of fairness in GNNs, as it shows that the model’s performance on different structural subgroups is influenced by their distance to the training set. Next, we extend the result of Theorem 1 to analyze the relation between the generalization of different structural subgroups and their distance to the training set. 4.2 Unfair Generalization Performance on Structural Subgroups Theorem 2 states that when comparing two subgroups, if the difference in their structural distance to the training set is larger than the threshold (δα), then GNNs exhibit unfair generalization performance with respect to these groups. Specifically, the theorem states that GNNs generalize better on subgroups that are structurally closer to the training set, which leads to accuracy disparities among different subgroups. Additionally, the theorem notes that the topology awareness of GNNs (measured by the distortion rate α) plays a key role in determining the strength of this accuracy disparity. This has important implications for the use of GNNs in critical applications, as it suggests that increased topology awareness in GNNs can lead to improved generalization performance but also a greater potential for unfairness in terms of accuracy disparities among different structural subgroups. As such, it is important to carefully consider the trade-offs between topology awareness and fairness when designing new GNNs and applying them to real-world problems. 4.3 Connection with Deep Learning Spline Theory Our theoretical findings reveal that when considering the performance of the model from the viewpoint of the embedding space, the training set creates a “valley” effect on generalization performance, with the training vertex at the bottom of the valley. Representations that are closer to the embedding of the training set perform better, while those farther away perform worse. This pattern has an interesting connection with deep learning spline theory [2]. This theory suggests that in DNNs, the learning process for classification divides the input/latent space into small circles (referred to as power diagrams), where representations within each circle are grouped, and representations closer to the centre of the circle are predicted with more confidence. Our findings suggest that in GNNs, the mapping of test data to these power diagrams is related to the structural similarity between test and training data in the graph. This connection provides further validation for our derived results. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

On the Topology Awareness and Generalization Performance of Graph Neural Networks: Main Results

Authors:
(1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk;
(2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction 2 Related Work 3 Framework 4 Main Results 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 11 Other Potential Applications 3 Framework In this section, we introduce our proposed framework, which aims to investigate the intricate relationship between topology awareness and generalization performance in GNNs. The main idea of the framework is to use metric distortion to characterize the topology awareness of GNNs. With this characterization, we are able to show the relation between topology awareness of GNNs and their generalization performance. 3.1 GNN Preliminaries Let G = (V, E, X) be an undirected graph where V = {1, 2, ..., n} is the set of n nodes and E ⊆ V × V is the set of edges. Let X = {x1, x2, ..., xn} denote the set of node attribute vectors corresponding to all the nodes in V. More specifically, xv ∈ X is an m-dimensional vector which captures the attribute values of node v ∈ V. In addition, each vertex v ∈ V is associated with a node label yv ∈ {1, ..., c}, where c is the total number of classes. For our analysis, we focus on the transductive node classification setting [47], where the graph G are observed prior to learning and labels are only available for a small subset of nodes V0 ⊂ V. Without loss of generality, we treat G as fixed throughout our analysis and the randomness comes from the labels y. Given a small set of the labelled nodes, V0 ⊂ V, the task of node-level semi-supervised learning is to learn a classifier f, such that yˆv = f(v), from a hypothesis space F and perform it on the remaining unlabeled nodes. Let L(., .) 7→ R+ be a bounded continuous loss function. We denote by RL(f, V) the generalization risk of the classifier f with respect to L and V and it is defined as follows, In our case, the hypothesis space is captured by a GNN and an additional read-out function. GNNs are commonly used as an embedding function and adopt a neighbourhood aggregation scheme, where the representation of a node is learnt iteratively by aggregating representations of its neighbours. After k iterations of aggregation, a node’s representation captures the information within its k-hop neighbourhood in the graph. Formally, the k-th layer of a GNN can be written as follows, It should be noted that the distance function in Def. 2 is not symmetrical, and Eq. 4 captures the maximal structural difference of vertices in V1 to V2. To provide further insight into the concept of a structural subgroup, let’s consider a scenario where V0 ⊂ V represents the training set and ds(.) represents the shortest-path distance. In this context, we can partition the remaining vertices V\V0 into different structural subgroups based on their shortest-path distance to the nearest vertex in the training set. Specifically, we can define Vk = {v ∈ V\V0|Ds(v, V0) = k}. However, it is important to emphasize that our analysis does not assume a specific partition scheme, and the framework remains adaptable to various scenarios. 3.3 Topology Awareness To characterize and formalize the topology awareness of GNNs, we introduce the concept of distortion, which measures the degree to which the graph structure is preserved in the embedding space. Mathematically, the distortion of a function between two general metric spaces is defined as, The definition above captures the maximum expansion and minimum contraction between the two metric spaces. In the extreme case when α = 1, the mapping γ is the so-called isometric mapping that perfectly preserves the distance between any two points in the space. In other words, for any two points x and y in the domain of the function, their distance d(x, y) in the original metric space is equal to the distance d ′ (γ(x), γ(y)) in the target metric space (up to a constant scaling factor). If the distortion rate of the mapping is close to 1, then this mapping is said to have a low distortion rate. In the context of GNNs, evaluating distortion involves a straightforward comparison of embeddings in the embedding domain with the structural measures on the graph, as demonstrated in the case study. If the distortion rate between the graph domain and the embedding domain is close to 1, the resulting embedding would largely preserve the corresponding structural information from the graph space, thereby indicating high topology awareness. Consequently, distortion can be employed to measure the topology awareness of GNNs concerning the corresponding structural properties (e.g., shortest-path distance) measured in the graph domain. Consequently, distortion can be employed to measure the topology awareness of GNNs concerning the corresponding structural properties (e.g., shortest-path distance) measured in the graph domain. This concept of distortion is agnostic to specific structural configurations, enabling the study of diverse graph structures using appropriate distance functions. Moreover, its estimation can be readily achieved by comparing the embedding with the original structural measure, as illustrated later in the case study This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Authors: Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Framework 3 Framework 4 Main Results 4 Main Results 5 A Case Study on Shortest-Path Distance 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 7 Proof of Theorem 1 8 Proof of Theorem 2 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 10 Additional Experiments Details and Results 11 Other Potential Applications 11 Other Potential Applications 3 Framework In this section, we introduce our proposed framework, which aims to investigate the intricate relationship between topology awareness and generalization performance in GNNs. The main idea of the framework is to use metric distortion to characterize the topology awareness of GNNs. With this characterization, we are able to show the relation between topology awareness of GNNs and their generalization performance. 3.1 GNN Preliminaries Let G = (V, E, X) be an undirected graph where V = {1, 2, ..., n} is the set of n nodes and E ⊆ V × V is the set of edges. Let X = {x1, x2, ..., xn} denote the set of node attribute vectors corresponding to all the nodes in V. More specifically, xv ∈ X is an m-dimensional vector which captures the attribute values of node v ∈ V. In addition, each vertex v ∈ V is associated with a node label yv ∈ {1, ..., c}, where c is the total number of classes. For our analysis, we focus on the transductive node classification setting [47], where the graph G are observed prior to learning and labels are only available for a small subset of nodes V0 ⊂ V. Without loss of generality, we treat G as fixed throughout our analysis and the randomness comes from the labels y. Given a small set of the labelled nodes, V0 ⊂ V, the task of node-level semi-supervised learning is to learn a classifier f, such that yˆv = f(v), from a hypothesis space F and perform it on the remaining unlabeled nodes. Let L(., .) 7→ R+ be a bounded continuous loss function. We denote by RL(f, V) the generalization risk of the classifier f with respect to L and V and it is defined as follows, In our case, the hypothesis space is captured by a GNN and an additional read-out function. GNNs are commonly used as an embedding function and adopt a neighbourhood aggregation scheme, where the representation of a node is learnt iteratively by aggregating representations of its neighbours. After k iterations of aggregation, a node’s representation captures the information within its k-hop neighbourhood in the graph. Formally, the k-th layer of a GNN can be written as follows, It should be noted that the distance function in Def. 2 is not symmetrical, and Eq. 4 captures the maximal structural difference of vertices in V1 to V2. To provide further insight into the concept of a structural subgroup, let’s consider a scenario where V0 ⊂ V represents the training set and ds(.) represents the shortest-path distance. In this context, we can partition the remaining vertices V\V0 into different structural subgroups based on their shortest-path distance to the nearest vertex in the training set. Specifically, we can define Vk = {v ∈ V\V0|Ds(v, V0) = k}. However, it is important to emphasize that our analysis does not assume a specific partition scheme, and the framework remains adaptable to various scenarios. 3.3 Topology Awareness To characterize and formalize the topology awareness of GNNs, we introduce the concept of distortion, which measures the degree to which the graph structure is preserved in the embedding space. Mathematically, the distortion of a function between two general metric spaces is defined as, The definition above captures the maximum expansion and minimum contraction between the two metric spaces. In the extreme case when α = 1, the mapping γ is the so-called isometric mapping that perfectly preserves the distance between any two points in the space. In other words, for any two points x and y in the domain of the function, their distance d(x, y) in the original metric space is equal to the distance d ′ (γ(x), γ(y)) in the target metric space (up to a constant scaling factor). If the distortion rate of the mapping is close to 1, then this mapping is said to have a low distortion rate. In the context of GNNs, evaluating distortion involves a straightforward comparison of embeddings in the embedding domain with the structural measures on the graph, as demonstrated in the case study. If the distortion rate between the graph domain and the embedding domain is close to 1, the resulting embedding would largely preserve the corresponding structural information from the graph space, thereby indicating high topology awareness. Consequently, distortion can be employed to measure the topology awareness of GNNs concerning the corresponding structural properties (e.g., shortest-path distance) measured in the graph domain. Consequently, distortion can be employed to measure the topology awareness of GNNs concerning the corresponding structural properties (e.g., shortest-path distance) measured in the graph domain. This concept of distortion is agnostic to specific structural configurations, enabling the study of diverse graph structures using appropriate distance functions. Moreover, its estimation can be readily achieved by comparing the embedding with the original structural measure, as illustrated later in the case study This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Framework for Analyzing Topology Awareness and Generalization in Graph Neural Networks

Authors:
(1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk;
(2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction 2 Related Work 3 Framework 4 Main Results 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 11 Other Potential Applications 2 Related Work 2.1 Topology Awareness of GNNs Modern GNN architectures take inspiration from the Weisfeiler-Lehman isomorphism test [12,19], which employs the graph structure to propagate information. Consequently, a significant portion of existing literature on GNNs concentrates on understanding their ability to differentiate various graph structures, known as the expressive power of GNNs [28–30, 44, 48, 50]. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties in the learned representation [29, 33, 48]. Nevertheless, the precise ramifications of heightened topology awareness on GNNs’ generalization performance remain shrouded in ambiguity. The intricate variations within structures across different domains and tasks further compound this complexity, underscoring the necessity for a unified framework—akin to the one proposed in this paper—to shed light on these intricacies. 2.2 Generalization of GNNs Relatively few studies focus on understanding the generalization performance of GNNs in node-level tasks. [3, 9, 18, 23, 25, 34, 39, 41, 45] extend the conventional statistical learning or information-theoretic framework to GNNs, providing different generalization bounds based on concepts such as Vapnik–Chervonenkis dimension, mutual information, and algorithm stability. In light of this, despite empirical indications highlighting the critical role of GNNs’ topology awareness in generalization performance [20,29,48], there remains limited understanding of the relationship between topology awareness of GNNs and their generalization. Our proposed framework addresses this gap by utilizing the distortion concept in the approximate metric embedding to connect the topology awareness of GNNs and their generalization performance. 2.3 Active Learning in Graph Active Learning (AL) seeks to enhance model performance within a limited labeling budget by identifying the most representative samples for annotation [32]. The AL methodology generally involves two phases: 1) selecting a small, preliminary subset of samples for labeling; 2) utilizing this labeled subset to iteratively discover the most informative sample for future labeling. The initial selection of an effective subset represents the cold-start problem in AL [11]. While AL has proven effective in settings with independent and identically distributed (i.i.d.) data, such as images and text, applying it to Graph Neural Networks (GNNs) poses unique challenges. These challenges stem from the necessity to account for node interdependencies and the computational complexity of iterative sample selection and retraining, especially when considering multi-hop neighbors. Recent AL methodologies for GNNs, such as AGE [5], FeatProp [42], and GraphPart [26], aim to overcome these obstacles by leveraging the graph structure and initial node features to boost efficiency and efficacy, although their success may significantly hinge on the quality of the initial features. In this work, we focus on the cold-start problem of graph active learning and introduce a novel approach (based on our analysis) that relies exclusively on graph structure, thus eliminating the reliance on feature quality at the onset of learning. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Authors: Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Framework 3 Framework 4 Main Results 4 Main Results 5 A Case Study on Shortest-Path Distance 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 7 Proof of Theorem 1 8 Proof of Theorem 2 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 10 Additional Experiments Details and Results 11 Other Potential Applications 11 Other Potential Applications 2 Related Work 2.1 Topology Awareness of GNNs Modern GNN architectures take inspiration from the Weisfeiler-Lehman isomorphism test [12,19], which employs the graph structure to propagate information. Consequently, a significant portion of existing literature on GNNs concentrates on understanding their ability to differentiate various graph structures, known as the expressive power of GNNs [28–30, 44, 48, 50]. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties in the learned representation [29, 33, 48]. Nevertheless, the precise ramifications of heightened topology awareness on GNNs’ generalization performance remain shrouded in ambiguity. The intricate variations within structures across different domains and tasks further compound this complexity, underscoring the necessity for a unified framework—akin to the one proposed in this paper—to shed light on these intricacies. 2.2 Generalization of GNNs Relatively few studies focus on understanding the generalization performance of GNNs in node-level tasks. [3, 9, 18, 23, 25, 34, 39, 41, 45] extend the conventional statistical learning or information-theoretic framework to GNNs, providing different generalization bounds based on concepts such as Vapnik–Chervonenkis dimension, mutual information, and algorithm stability. In light of this, despite empirical indications highlighting the critical role of GNNs’ topology awareness in generalization performance [20,29,48], there remains limited understanding of the relationship between topology awareness of GNNs and their generalization. Our proposed framework addresses this gap by utilizing the distortion concept in the approximate metric embedding to connect the topology awareness of GNNs and their generalization performance. 2.3 Active Learning in Graph Active Learning (AL) seeks to enhance model performance within a limited labeling budget by identifying the most representative samples for annotation [32]. The AL methodology generally involves two phases: 1) selecting a small, preliminary subset of samples for labeling; 2) utilizing this labeled subset to iteratively discover the most informative sample for future labeling. The initial selection of an effective subset represents the cold-start problem in AL [11]. While AL has proven effective in settings with independent and identically distributed (i.i.d.) data, such as images and text, applying it to Graph Neural Networks (GNNs) poses unique challenges. These challenges stem from the necessity to account for node interdependencies and the computational complexity of iterative sample selection and retraining, especially when considering multi-hop neighbors. Recent AL methodologies for GNNs, such as AGE [5], FeatProp [42], and GraphPart [26], aim to overcome these obstacles by leveraging the graph structure and initial node features to boost efficiency and efficacy, although their success may significantly hinge on the quality of the initial features. In this work, we focus on the cold-start problem of graph active learning and introduce a novel approach (based on our analysis) that relies exclusively on graph structure, thus eliminating the reliance on feature quality at the onset of learning. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Exploring Topology Awareness, Generalization, and Active Learning in Graph Neural Networks

Authors:
(1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk;
(2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction 2 Related Work 3 Framework 4 Main Results 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 11 Other Potential Applications Abstract Many computer vision and machine learning problems are modelled as learning tasks on graphs, where graph neural networks (GNNs) have emerged as a dominant tool for learning representations of graphstructured data. A key feature of GNNs is their use of graph structures as input, enabling them to exploit the graphs’ inherent topological properties—known as the topology awareness of GNNs. Despite the empirical successes of GNNs, the influence of topology awareness on generalization performance remains unexplored, particularly for node-level tasks that diverge from the assumption of data being independent and identically distributed (I.I.D.). The precise definition and characterization of the topology awareness of GNNs, especially concerning different topological features, are still unclear. This paper introduces a comprehensive framework to characterize the topology awareness of GNNs across any topological feature. Using this framework, we investigate the effects of topology awareness on GNN generalization performance. Contrary to the prevailing belief that enhancing the topology awareness of GNNs is always advantageous, our analysis reveals a critical insight: improving the topology awareness of GNNs may inadvertently lead to unfair generalization across structural groups, which might not be desired in some scenarios. Additionally, we conduct a case study using the intrinsic graph metric, the shortest-path distance, on various benchmark datasets. The empirical results of this case study confirm our theoretical insights. Moreover, we demonstrate the practical applicability of our framework by using it to tackle the cold start problem in graph active learning. 1 Introduction Many problems in computer vision and machine learning are modeled as learning tasks on graphs. For example, in semantic segmentation, graphs model the relationships between different image regions, enhancing accuracy and context-aware segmentation. Graph neural networks (GNNs) have emerged as a dominant class of machine learning models specifically designed for learning representations of graph-structured data. They have demonstrated considerable success in addressing a wide range of graph-related problems in various domains such as chemistry [10], biology [37], social networking [6, 22], scene graph generation [46, 51] and visual relationship detection [24,43,49]. A defining characteristic of GNNs is their use of a spatial approach through message passing on the graph structure for feature aggregation. This enables GNNs to preserve structural information or dependencies (referred to as topology awareness) from the underlying graph structure, allowing them to be highly effective in tasks such as node classification. Fig. 1 illustrates the overall learning process of GNNs. Despite their practicality and potential, there remains a lack of theoretical understanding about GNNs, particularly in the semi-supervised node classification setting where the dependencies among the data differ significantly from other machine learning models [25]. In this setting, the goal is to leverage relations, as captured by the graph structure, among the data and a small set of labelled nodes to predict labels for the remaining nodes. Most of the existing theoretical studies of GNNs have focused on the connection between the message-passing mechanism of GNNs and the Weisfeiler-Lehman isomorphism test [19], aiming to understand the ability of GNNs to differentiate different graph structures in the learned representations, known as the expressive power of GNNs. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties in the learned representation [29, 33, 48]. However, as GNNs become more reliant on and sensitive (aware) of graph structure as input, they may exhibit different generalization performances towards certain structural subgroups (distinct data subsets grouped by structural similarity to the training set) within the data. The quantification of GNN generalization across distinct structural subgroups is termed structural subgroup generalization [25]. Such considerations are vital in GNN application and development. For instance, within protein-protein interaction networks, these structural subgroups could represent different molecular complexes, influencing the accuracy of interaction predictions. Similarly, understanding how the topology awareness of GNNs influences generalization is essential when devising sampling strategies for training. The extent to which the generalization performance of GNNs is influenced by specific structural features of graph data is critical in deciding the composition of training datasets. Despite its importance, an understanding of the relationship between the topology awareness of GNNs and its structural subgroup generalization is still lacking. Furthermore, characterizing the topology awareness of GNNs poses a challenge, especially considering that different domains and tasks may prioritize distinct structural aspects. Therefore, a versatile framework is needed to assess the topology awareness of GNNs in relation to various structures. To address this gap, in this paper, we propose a novel framework based on approximate metric embedding to study the relationship between structural subgroup generalization and topology awareness of GNNs in the context of semisupervised node classification. The proposed framework allows for the investigation of the structural subgroup generalization of GNNs with respect to different structural subgroups. More concretely, the main contributions of this work are summarized as follows. 1. We propose a novel, structure-agnostic framework using approximate metric embedding to examine the interplay between GNNs’ structural subgroup generalization and topology awareness. This framework is versatile, accommodating various structural measures like shortest-path distance, and requires only the corresponding structural measure. Its simplicity in estimating key factors makes it applicable and generalizable to a wide range of scenarios. 2. Through formal analysis within our framework, we establish a clear link between GNN topology awareness and their generalization performance (Theorem 1). We also demonstrate that while enhanced topology awareness boosts GNN expressiveness, it can result in uneven generalization performance, favouring subgroups more structurally similar to the training set (Theorem 2). Such structural property can be harmful (causing unfairness issues) or useful (informing design decisions) depending on the scenario. This challenges the prevailing belief that increased topology awareness universally benefits GNNs [29, 33, 48], emphasizing the importance of considering the relationship between topology awareness and generalization performance. 3. We validate our framework through a case study on shortest-path distance, highlighting its practicality and relevance. The results corroborate our theoretical findings, showing that GNNs with heightened awareness of shortestpath distances excel in classifying vertex groups closer to the training set. Moreover, we demonstrate how our findings can be applied to mitigate the cold start problem in graph active learning [11,15], highlighting the practical implications of our framework and results. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Authors: Authors: (1) Junwei Su, Department of Computer Science, the University of Hong Kong and jwsu@cs.hku.hk; (2) Chuan Wu, Department of Computer Science, the University of Hong Kong and cwu@cs.hku.hk. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Framework 3 Framework 4 Main Results 4 Main Results 5 A Case Study on Shortest-Path Distance 5 A Case Study on Shortest-Path Distance 6 Conclusion and Discussion, and References 6 Conclusion and Discussion, and References 7 Proof of Theorem 1 7 Proof of Theorem 1 8 Proof of Theorem 2 8 Proof of Theorem 2 9 Procedure for Solving Eq. (6) 9 Procedure for Solving Eq. (6) 10 Additional Experiments Details and Results 10 Additional Experiments Details and Results 11 Other Potential Applications 11 Other Potential Applications Abstract Many computer vision and machine learning problems are modelled as learning tasks on graphs, where graph neural networks (GNNs) have emerged as a dominant tool for learning representations of graphstructured data. A key feature of GNNs is their use of graph structures as input, enabling them to exploit the graphs’ inherent topological properties—known as the topology awareness of GNNs. Despite the empirical successes of GNNs, the influence of topology awareness on generalization performance remains unexplored, particularly for node-level tasks that diverge from the assumption of data being independent and identically distributed (I.I.D.). The precise definition and characterization of the topology awareness of GNNs, especially concerning different topological features, are still unclear. This paper introduces a comprehensive framework to characterize the topology awareness of GNNs across any topological feature. Using this framework, we investigate the effects of topology awareness on GNN generalization performance. Contrary to the prevailing belief that enhancing the topology awareness of GNNs is always advantageous, our analysis reveals a critical insight: improving the topology awareness of GNNs may inadvertently lead to unfair generalization across structural groups, which might not be desired in some scenarios. Additionally, we conduct a case study using the intrinsic graph metric, the shortest-path distance, on various benchmark datasets. The empirical results of this case study confirm our theoretical insights. Moreover, we demonstrate the practical applicability of our framework by using it to tackle the cold start problem in graph active learning. 1 Introduction Many problems in computer vision and machine learning are modeled as learning tasks on graphs. For example, in semantic segmentation, graphs model the relationships between different image regions, enhancing accuracy and context-aware segmentation. Graph neural networks (GNNs) have emerged as a dominant class of machine learning models specifically designed for learning representations of graph-structured data. They have demonstrated considerable success in addressing a wide range of graph-related problems in various domains such as chemistry [10], biology [37], social networking [6, 22], scene graph generation [46, 51] and visual relationship detection [24,43,49]. A defining characteristic of GNNs is their use of a spatial approach through message passing on the graph structure for feature aggregation. This enables GNNs to preserve structural information or dependencies (referred to as topology awareness) from the underlying graph structure, allowing them to be highly effective in tasks such as node classification. Fig. 1 illustrates the overall learning process of GNNs. Despite their practicality and potential, there remains a lack of theoretical understanding about GNNs, particularly in the semi-supervised node classification setting where the dependencies among the data differ significantly from other machine learning models [25]. In this setting, the goal is to leverage relations, as captured by the graph structure, among the data and a small set of labelled nodes to predict labels for the remaining nodes. Most of the existing theoretical studies of GNNs have focused on the connection between the message-passing mechanism of GNNs and the Weisfeiler-Lehman isomorphism test [19], aiming to understand the ability of GNNs to differentiate different graph structures in the learned representations, known as the expressive power of GNNs. Inspired by the expressiveness studies, it is commonly believed that increasing topology awareness is universally beneficial and many studies focus on enabling GNNs to preserve more structural properties in the learned representation [29, 33, 48]. However, as GNNs become more reliant on and sensitive (aware) of graph structure as input, they may exhibit different generalization performances towards certain structural subgroups (distinct data subsets grouped by structural similarity to the training set) within the data. The quantification of GNN generalization across distinct structural subgroups is termed structural subgroup generalization [25]. Such considerations are vital in GNN application and development. For instance, within protein-protein interaction networks, these structural subgroups could represent different molecular complexes, influencing the accuracy of interaction predictions. Similarly, understanding how the topology awareness of GNNs influences generalization is essential when devising sampling strategies for training. The extent to which the generalization performance of GNNs is influenced by specific structural features of graph data is critical in deciding the composition of training datasets. Despite its importance, an understanding of the relationship between the topology awareness of GNNs and its structural subgroup generalization is still lacking. Furthermore, characterizing the topology awareness of GNNs poses a challenge, especially considering that different domains and tasks may prioritize distinct structural aspects. Therefore, a versatile framework is needed to assess the topology awareness of GNNs in relation to various structures. To address this gap, in this paper, we propose a novel framework based on approximate metric embedding to study the relationship between structural subgroup generalization and topology awareness of GNNs in the context of semisupervised node classification. The proposed framework allows for the investigation of the structural subgroup generalization of GNNs with respect to different structural subgroups. More concretely, the main contributions of this work are summarized as follows. 1. We propose a novel, structure-agnostic framework using approximate metric embedding to examine the interplay between GNNs’ structural subgroup generalization and topology awareness. This framework is versatile, accommodating various structural measures like shortest-path distance, and requires only the corresponding structural measure. Its simplicity in estimating key factors makes it applicable and generalizable to a wide range of scenarios. 2. Through formal analysis within our framework, we establish a clear link between GNN topology awareness and their generalization performance (Theorem 1). We also demonstrate that while enhanced topology awareness boosts GNN expressiveness, it can result in uneven generalization performance, favouring subgroups more structurally similar to the training set (Theorem 2). Such structural property can be harmful (causing unfairness issues) or useful (informing design decisions) depending on the scenario. This challenges the prevailing belief that increased topology awareness universally benefits GNNs [29, 33, 48], emphasizing the importance of considering the relationship between topology awareness and generalization performance. 3. We validate our framework through a case study on shortest-path distance, highlighting its practicality and relevance. The results corroborate our theoretical findings, showing that GNNs with heightened awareness of shortestpath distances excel in classifying vertex groups closer to the training set. Moreover, we demonstrate how our findings can be applied to mitigate the cold start problem in graph active learning [11,15], highlighting the practical implications of our framework and results. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Understanding Topology Awareness in Graph Neural Networks

Comprensión de la conciencia topológica en redes neuronales gráficas: impactos en la generalización y la estructura

Konprann konsyantizasyon Topoloji nan rezo neral graf: enpak sou jeneralizasyon ak estriktirèl 

グラフニューラルネットワークにおけるトポロジー認識の理解：一般化と構造への影響

График мэдрэлийн сүлжээн дэх топологийн мэдлэгийг ойлгох нь: ерөнхий ойлголт ба бүтцийн байдалд үзүүлэх нөлөө 

ग्राफ न्यूरल नेटवर्कमा टोपोलोजी जागरूकता बुझ्दै: सामान्यीकरण र संरचनात्मकमा प्रभावहरू 

Go kwešiša Temogo ya Topology ka go Dinetweke tša Neural tša Kerafo: Ditlamorago go Kakaretšo le Sebopego 

په ګراف عصبي شبکو کې د توپولوژي پوهاوی پوهیدل: په عمومي کولو او جوړښت باندې اغیزې 

Kunzwisisa Kuzivikanwa kweTopology muGraph Neural Networks: Mhedzisiro paGeneralization uye Chimiro. 

การทำความเข้าใจการรับรู้เกี่ยวกับโทโพโลยีในเครือข่ายประสาทเทียมแบบกราฟ: ผลกระทบต่อการสรุปทั่วไปและโครงสร้าง

Pag-unawa sa Topology Awareness sa Graph Neural Networks: Mga Epekto sa Generalization at Structural 

گراف نیورل نیٹ ورکس میں ٹوپولوجی بیداری کو سمجھنا: عمومی اور ساختی اثرات پر 

Grafik neyron tarmoqlarida topologiyadan xabardorlikni tushunish: umumlashtirish va strukturaga ta'siri 

Authors:
(1) Seokil Ham, KAIST;
(2) Jungwuk Park, KAIST;
(3) Dong-Jun Han, Purdue University;
(4) Jaekyun Moon, KAIST. Table of Links Abstract and 1. Introduction 2. Related Works 3. Proposed NEO-KD Algorithm and 3.1 Problem Setup: Adversarial Training in Multi-Exit Networks 3.2 Algorithm Description 4. Experiments and 4.1 Experimental Setup 4.2. Main Experimental Results 4.3. Ablation Studies and Discussions 5. Conclusion, Acknowledgement and References A. Experiment Details B. Clean Test Accuracy and C. Adversarial Training via Average Attack D. Hyperparameter Tuning E. Discussions on Performance Degradation at Later Exits F. Comparison with Recent Defense Methods for Single-Exit Networks G. Comparison with SKD and ARD and H. Implementations of Stronger Attacker Algorithms Abstract While multi-exit neural networks are regarded as a promising solution for making efficient inference via early exits, combating adversarial attacks remains a challenging problem. In multi-exit networks, due to the high dependency among different submodels, an adversarial example targeting a specific exit not only degrades the performance of the target exit but also reduces the performance of all other exits concurrently. This makes multi-exit networks highly vulnerable to simple adversarial attacks. In this paper, we propose NEO-KD, a knowledge-distillation-based adversarial training strategy that tackles this fundamental challenge based on two key contributions. NEO-KD first resorts to neighbor knowledge distillation to guide the output of the adversarial examples to tend to the ensemble outputs of neighbor exits of clean data. NEO-KD also employs exit-wise orthogonal knowledge distillation for reducing adversarial transferability across different submodels. The result is a significantly improved robustness against adversarial attacks. Experimental results on various datasets/models show that our method achieves the best adversarial accuracy with reduced computation budgets, compared to the baselines relying on existing adversarial training or knowledge distillation techniques for multi-exit networks. 1 Introduction Multi-exit neural networks are receiving significant attention [9, 13, 26, 27, 28, 32] for their ability to make dynamic predictions in resource-constrained applications. Instead of making predictions at the final output of the full model, a faster prediction can be made at an earlier exit depending on the current time budget or computing budget. In this sense, a multi-exit network can be viewed as an architecture having multiple submodels, where each submodel consists of parameters from the input of the model to the output of a specific exit. These submodels are highly correlated as they share some model parameters. It is also well-known that the performance of all submodels can be improved by distilling the knowledge of the last exit to other exits, i.e., via self-distillation [15, 20, 24, 27]. There have also been efforts to address the adversarial attack issues in the context of multi-exit networks [3, 12]. Providing robustness against adversarial attacks is especially challenging in multi-exit networks: since different submodels have high correlations by sharing parameters, an adversarial example targeting a specific exit can significantly degrade the performance of other submodels. In other words, an adversarial example can have strong adversarial transferability across different submodels, making the model highly vulnerable to simple adversarial attacks (e.g., an adversarial attack targeting a single exit). Motivation. Only a few prior works have focused on adversarial defense strategies for multi-exit networks [3, 12]. The authors of [12] focused on generating adversarial examples tailored to multiexit networks (e.g., generate samples via max-average attack), and trained the model to minimize the sum of clean and adversarial losses of all exits. Given the adversarial example constructed in [12], the authors of [3] proposed a regularization term to reduce the weights of the classifier at each exit during training. However, existing adversarial defense strategies [3, 12] do not directly handle the high correlations among different submodels, resulting in high adversarial transferability and limited robustness in multi-exit networks. To tackle this difficulty, we take a knowledge-distillation-based approach in a fashion orthogonal to prior works [3, 12]. Some previous studies [8, 23, 33, 34] have shown that knowledge distillation can be utilized for improving the robustness of the model in conventional single-exit networks. However, although there are extensive existing works on self-distillation for training multi-exit networks using clean data [15, 20, 24, 27], it is currently unknown how distillation techniques should be utilized for adversarial training of multi-exit networks. Moreover, when the existing distillation-based schemes are applied to multi-exit networks, the dependencies among submodels become higher since the same output (e.g., the knowledge of the last exit) is distilled to all sub-models. Motivated by these limitations, we pose the following questions: How can we take advantage of knowledge-distillation to improve adversarial robustness of multi-exit networks? At the same time, how can we reduce adversarial transferability across different submodels in multi-exit networks? Main contributions. To handle these questions, we propose NEO-KD, a knowledge-distillation-based adversarial training strategy highly tailored to robust multi-exit neural networks. Our solution is two-pronged: neighbor knowledge distillation and exit-wise orthogonal knowledge distillation. • Given a specific exit, the first part of our solution, neighbor knowledge distillation (NKD), distills the ensembled prediction of neighbor exits of clean data to the prediction of the adversarial example at the corresponding exit, as shown in Figure 1a. This method guides the output of adversarial examples to follow the outputs of clean data, improving robustness against adversarial attacks. By ensembling the neighbor predictions of clean data before distillation, NKD provides higher quality features to the corresponding exits compared to the scheme distilling with only one exit in the same position. • The second focus of our solution, exit-wise orthogonal knowledge distillation (EOKD), mainly aims at reducing adversarial transferability across different submodels. This part is another unique contribution of our work compared to existing methods on robust multi-exit networks [3, 12] (that suffer from high adversarial transferability) or self-distillation-based multi-exit networks [15, 20, 24, 27] (that further increase adversarial transferability). In our EOKD, the output of clean data at the i-th exit is distilled to the output of the adversarial sample at the i-th exit, in an exit-wise manner. During this exit-wise distillation process, we encourage the non-ground-truth predictions of individual exits to be mutually orthogonal, by providing orthogonal soft labels to each exit as described in Figure 1b. By weakening the dependencies among different exit outputs, EOKD reduces the adversarial transferability across all submodels in the network, which leads to an improved robustness against adversarial attacks. The NKD and EOKD components of our architectural solution work together to reduce adversarial transferability across different submodels in the network while correctly guiding the predictions of the adversarial examples at each exit. Experimental results on various datasets show that the proposed strategy achieves the best adversarial accuracy with reduced computation budgets, compared to existing adversarial training methods for multi-exit networks. Our solution is a plug-and-play method, which can be used in conjunction with existing training strategies tailored to multi-exit networks. This paper is available on arxiv under CC 4.0 license. Authors: (1) Seokil Ham, KAIST; (2) Jungwuk Park, KAIST; (3) Dong-Jun Han, Purdue University; (4) Jaekyun Moon, KAIST. Authors: Authors: (1) Seokil Ham, KAIST; (2) Jungwuk Park, KAIST; (3) Dong-Jun Han, Purdue University; (4) Jaekyun Moon, KAIST. Table of Links Abstract and 1. Introduction Abstract and 1. Introduction 2. Related Works 2. Related Works 3. Proposed NEO-KD Algorithm and 3.1 Problem Setup: Adversarial Training in Multi-Exit Networks 3. Proposed NEO-KD Algorithm and 3.1 Problem Setup: Adversarial Training in Multi-Exit Networks 3.2 Algorithm Description 3.2 Algorithm Description 4. Experiments and 4.1 Experimental Setup 4. Experiments and 4.1 Experimental Setup 4.2. Main Experimental Results 4.2. Main Experimental Results 4.3. Ablation Studies and Discussions 4.3. Ablation Studies and Discussions 5. Conclusion, Acknowledgement and References 5. Conclusion, Acknowledgement and References A. Experiment Details A. Experiment Details B. Clean Test Accuracy and C. Adversarial Training via Average Attack B. Clean Test Accuracy and C. Adversarial Training via Average Attack D. Hyperparameter Tuning D. Hyperparameter Tuning E. Discussions on Performance Degradation at Later Exits E. Discussions on Performance Degradation at Later Exits F. Comparison with Recent Defense Methods for Single-Exit Networks F. Comparison with Recent Defense Methods for Single-Exit Networks G. Comparison with SKD and ARD and H. Implementations of Stronger Attacker Algorithms G. Comparison with SKD and ARD and H. Implementations of Stronger Attacker Algorithms Abstract While multi-exit neural networks are regarded as a promising solution for making efficient inference via early exits, combating adversarial attacks remains a challenging problem. In multi-exit networks, due to the high dependency among different submodels, an adversarial example targeting a specific exit not only degrades the performance of the target exit but also reduces the performance of all other exits concurrently. This makes multi-exit networks highly vulnerable to simple adversarial attacks. In this paper, we propose NEO-KD, a knowledge-distillation-based adversarial training strategy that tackles this fundamental challenge based on two key contributions. NEO-KD first resorts to neighbor knowledge distillation to guide the output of the adversarial examples to tend to the ensemble outputs of neighbor exits of clean data. NEO-KD also employs exit-wise orthogonal knowledge distillation for reducing adversarial transferability across different submodels. The result is a significantly improved robustness against adversarial attacks. Experimental results on various datasets/models show that our method achieves the best adversarial accuracy with reduced computation budgets, compared to the baselines relying on existing adversarial training or knowledge distillation techniques for multi-exit networks. 1 Introduction Multi-exit neural networks are receiving significant attention [9, 13, 26, 27, 28, 32] for their ability to make dynamic predictions in resource-constrained applications. Instead of making predictions at the final output of the full model, a faster prediction can be made at an earlier exit depending on the current time budget or computing budget. In this sense, a multi-exit network can be viewed as an architecture having multiple submodels, where each submodel consists of parameters from the input of the model to the output of a specific exit. These submodels are highly correlated as they share some model parameters. It is also well-known that the performance of all submodels can be improved by distilling the knowledge of the last exit to other exits, i.e., via self-distillation [15, 20, 24, 27]. There have also been efforts to address the adversarial attack issues in the context of multi-exit networks [3, 12]. Providing robustness against adversarial attacks is especially challenging in multi-exit networks: since different submodels have high correlations by sharing parameters, an adversarial example targeting a specific exit can significantly degrade the performance of other submodels. In other words, an adversarial example can have strong adversarial transferability across different submodels, making the model highly vulnerable to simple adversarial attacks (e.g., an adversarial attack targeting a single exit). Motivation . Only a few prior works have focused on adversarial defense strategies for multi-exit networks [3, 12]. The authors of [12] focused on generating adversarial examples tailored to multiexit networks (e.g., generate samples via max-average attack), and trained the model to minimize the sum of clean and adversarial losses of all exits. Given the adversarial example constructed in [12], the authors of [3] proposed a regularization term to reduce the weights of the classifier at each exit during training. However, existing adversarial defense strategies [3, 12] do not directly handle the high correlations among different submodels, resulting in high adversarial transferability and limited robustness in multi-exit networks. To tackle this difficulty, we take a knowledge-distillation-based approach in a fashion orthogonal to prior works [3, 12]. Some previous studies [8, 23, 33, 34] have shown that knowledge distillation can be utilized for improving the robustness of the model in conventional single-exit networks. However, although there are extensive existing works on self-distillation for training multi-exit networks using clean data [15, 20, 24, 27], it is currently unknown how distillation techniques should be utilized for adversarial training of multi-exit networks. Moreover, when the existing distillation-based schemes are applied to multi-exit networks, the dependencies among submodels become higher since the same output (e.g., the knowledge of the last exit) is distilled to all sub-models. Motivated by these limitations, we pose the following questions: How can we take advantage of knowledge-distillation to improve adversarial robustness of multi-exit networks? At the same time, how can we reduce adversarial transferability across different submodels in multi-exit networks? Motivation How can we take advantage of knowledge-distillation to improve adversarial robustness of multi-exit networks? At the same time, how can we reduce adversarial transferability across different submodels in multi-exit networks? Main contributions. To handle these questions, we propose NEO-KD, a knowledge-distillation-based adversarial training strategy highly tailored to robust multi-exit neural networks. Our solution is two-pronged: neighbor knowledge distillation and exit-wise orthogonal knowledge distillation. Main contributions. • Given a specific exit, the first part of our solution, neighbor knowledge distillation (NKD), distills the ensembled prediction of neighbor exits of clean data to the prediction of the adversarial example at the corresponding exit, as shown in Figure 1a. This method guides the output of adversarial examples to follow the outputs of clean data, improving robustness against adversarial attacks. By ensembling the neighbor predictions of clean data before distillation, NKD provides higher quality features to the corresponding exits compared to the scheme distilling with only one exit in the same position. • The second focus of our solution, exit-wise orthogonal knowledge distillation (EOKD), mainly aims at reducing adversarial transferability across different submodels. This part is another unique contribution of our work compared to existing methods on robust multi-exit networks [3, 12] (that suffer from high adversarial transferability) or self-distillation-based multi-exit networks [15, 20, 24, 27] (that further increase adversarial transferability). In our EOKD, the output of clean data at the i-th exit is distilled to the output of the adversarial sample at the i-th exit, in an exit-wise manner. During this exit-wise distillation process, we encourage the non-ground-truth predictions of individual exits to be mutually orthogonal, by providing orthogonal soft labels to each exit as described in Figure 1b. By weakening the dependencies among different exit outputs, EOKD reduces the adversarial transferability across all submodels in the network, which leads to an improved robustness against adversarial attacks. The NKD and EOKD components of our architectural solution work together to reduce adversarial transferability across different submodels in the network while correctly guiding the predictions of the adversarial examples at each exit. Experimental results on various datasets show that the proposed strategy achieves the best adversarial accuracy with reduced computation budgets, compared to existing adversarial training methods for multi-exit networks. Our solution is a plug-and-play method, which can be used in conjunction with existing training strategies tailored to multi-exit networks. This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

Knowledge-Distillation-Based Adversarial Training for Robust Multi-Exit Neural Networks

التدريب التنافسي القائم على تقطير المعرفة للشبكات العصبية القوية متعددة المخارج

Yatiña-Destilación-Basado en Entrenamiento Adversario ukaxa Redes Neuronales Multi-Salida Robustas ukanakataki

Möhkəm Çox Çıxışlı Neyron Şəbəkələri üçün Bilik Distilləsinə əsaslanan rəqib təlimi

Adversariální školení pro robustní multi-exitové neuronové sítě založené na znalostní destilaci

Entrenamiento adversarial basado en destilación de conocimiento para redes neuronales robustas de múltiples salidas

Tietoon perustuvaa kontradiktorista koulutusta vankille monilähtöisille hermoverkoille

Formazione avversaria basata sulla distillazione della conoscenza per reti neurali multi-uscita robuste

堅牢なマルチエグジットニューラルネットワークのための知識蒸留ベースの敵対的トレーニング

Мықты көп шығу нейрондық желілері үшін білімге негізделген бәсекелестік тренинг

강력한 다중 출구 신경망을 위한 지식 증류 기반 적대적 훈련

Treinamento Adversarial Baseado em Destilação de Conhecimento para Redes Neurais Multi-Saídas Robustas

Antrenament adversar bazat pe distilare de cunoștințe pentru rețele neuronale robuste cu mai multe ieșiri

Cybersecurity threats are evolving at an alarming rate. From sophisticated malware to state-sponsored attacks, businesses worldwide face a battle as they attempt to safeguard their data. According to a report by Cybersecurity Ventures, the cost of cybercrime reached $8 trillion in 2023 alone and is projected to rise to $10.5 trillion by 2025. Businesses need robust security measures and future-proof strategies to prevent time consuming and costly security breaches. Unfortunately, attackers can exploit vulnerabilities from any point within a system. From the base hardware level to man-in-the-middle attacks and even social engineering, organisations still need a sufficiently intelligent threat prevention protocol. Security vendor bottlenecks leave them vulnerable to weaknesses. In these instances, businesses are over-reliant on a single solution or provider. Brander Group, a leading IT and network infrastructure consulting firm, tackles this issue with its vendor-neutral methodology. Offering a variety of network security & vulnerabilities solutions, Brander Group protects clients with multiple layers of redundancy so their network is avoids being at risk with a single point of failure. The Rising Tide of Network Security Threats In recent years, there has been a dramatic surge in sophisticated cyber attacks, with malware, ransomware, and advanced persistent threats (APTs) becoming increasingly prevalent. "The cybersecurity world has evolved at a dangerous pace," says Jake Brander, founder and CEO of Brander Group. "What worked yesterday may be obsolete today. That's why a flexible, vendor-neutral strategy is critical for long-term security." Vendor-Neutral Strategy: A Shield Against Vulnerabilities Brander Group specialises in vendor-neutral methods which are integral to protecting our clients. The company offers a diverse array of services & solutions, made possible by its partnerships with over 600 service providers globally. Brander Group provides each client with multiple options, fully managed deployments and support custom-tailored to their individual requirements and architectures. This strategy allows businesses to avoid the dangers of vendor lock-in, where dependence on a single provider can make it difficult to quickly adapt to new threats. Instead, Brander Group's clients benefit from a best-of-breed strategy, combining services from multiple providers to create a stronger security posture. "The idea isn’t just to avoid lock-ins. It's about creating a security ecosystem greater than the sum of its parts. We can mix and match solutions that are designed specifically for each client's needs," Brander explains. Brander Group's Expertise: The Human Element in Cybersecurity While the vendor-neutral strategy provides the framework, Brander Group's team of experts brings it to life. The company's provider-neutral engineers and consultants possess a deep understanding across a wide range of security protocols, allowing them to effectively and accurately wield each solution. "Technology is only as effective as the people implementing it. Moreover, the people implementing it must understand what is needed to effectively deliver each solution ." states Brander. Impenetrability Through Neutrality Brander Group’s services extend beyond simple security provision and implementation. Its team works closely with vendors and clients to develop comprehensive security strategies, conduct regular assessments, and provide ongoing support and training. With any system or device susceptible to several types of new age attacks, businesses worldwide are increasingly understanding the value of having a vendor-neutral network strategy. As Jake Brander aptly puts it, “There is no such thing as a truly secure system. But, with multiple layers of security, we can make it virtually impossible for any one security hole to be exploited. This is the true advantage of working with Brander Group – impenetrability through neutrality.” This piece was published as part of HackerNoon’s Business Blogging Program. For more details, please click here. Cybersecurity threats are evolving at an alarming rate. From sophisticated malware to state-sponsored attacks, businesses worldwide face a battle as they attempt to safeguard their data. According to a report by Cybersecurity Ventures, the cost of cybercrime reached $8 trillion in 2023 alone and is projected to rise to $10.5 trillion by 2025. Businesses need robust security measures and future-proof strategies to prevent time consuming and costly security breaches. Unfortunately, attackers can exploit vulnerabilities from any point within a system. From the base hardware level to man-in-the-middle attacks and even social engineering, organisations still need a sufficiently intelligent threat prevention protocol. Security vendor bottlenecks leave them vulnerable to weaknesses. In these instances, businesses are over-reliant on a single solution or provider. Brander Group, a leading IT and network infrastructure consulting firm, tackles this issue with its vendor-neutral methodology. Offering a variety of network security & vulnerabilities solutions, Brander Group protects clients with multiple layers of redundancy so their network is avoids being at risk with a single point of failure. network security The Rising Tide of Network Security Threats The Rising Tide of Network Security Threats In recent years, there has been a dramatic surge in sophisticated cyber attacks, with malware, ransomware, and advanced persistent threats (APTs) becoming increasingly prevalent. "The cybersecurity world has evolved at a dangerous pace," says Jake Brander, founder and CEO of Brander Group. "What worked yesterday may be obsolete today. That's why a flexible, vendor-neutral strategy is critical for long-term security." Vendor-Neutral Strategy: A Shield Against Vulnerabilities Vendor-Neutral Strategy: A Shield Against Vulnerabilities Brander Group specialises in vendor-neutral methods which are integral to protecting our clients. The company offers a diverse array of services & solutions, made possible by its partnerships with over 600 service providers globally. Brander Group provides each client with multiple options, fully managed deployments and support custom-tailored to their individual requirements and architectures. This strategy allows businesses to avoid the dangers of vendor lock-in, where dependence on a single provider can make it difficult to quickly adapt to new threats. Instead, Brander Group's clients benefit from a best-of-breed strategy, combining services from multiple providers to create a stronger security posture. " The idea isn’t just to avoid lock-ins. It's about creating a security ecosystem greater than the sum of its parts. We can mix and match solutions that are designed specifically for each client's needs ," Brander explains. The idea isn’t just to avoid lock-ins. It's about creating a security ecosystem greater than the sum of its parts. We can mix and match solutions that are designed specifically for each client's needs Brander Group's Expertise: The Human Element in Cybersecurity Brander Group's Expertise: The Human Element in Cybersecurity While the vendor-neutral strategy provides the framework, Brander Group's team of experts brings it to life. The company's provider-neutral engineers and consultants possess a deep understanding across a wide range of security protocols, allowing them to effectively and accurately wield each solution. " Technology is only as effective as the people implementing it. Moreover, the people implementing it must understand what is needed to effectively deliver each solution . " states Brander. Technology is only as effective as the people implementing it. Moreover, the people implementing it must understand what is needed to effectively deliver each solution . Impenetrability Through Neutrality Impenetrability Through Neutrality Brander Group’s services extend beyond simple security provision and implementation. Its team works closely with vendors and clients to develop comprehensive security strategies, conduct regular assessments, and provide ongoing support and training. With any system or device susceptible to several types of new age attacks, businesses worldwide are increasingly understanding the value of having a vendor-neutral network strategy. As Jake Brander aptly puts it, “ There is no such thing as a truly secure system. But, with multiple layers of security, we can make it virtually impossible for any one security hole to be exploited. This is the true advantage of working with Brander Group – impenetrability through neutrality. ” There is no such thing as a truly secure system. But, with multiple layers of security, we can make it virtually impossible for any one security hole to be exploited. This is the true advantage of working with Brander Group – impenetrability through neutrality. This piece was published as part of HackerNoon’s Business Blogging Program. For more details, please click here. This piece was published as part of HackerNoon’s Business Blogging Program. For more details, please click here. This piece was published as part of HackerNoon’s Business Blogging Program. For more details, please click here. This piece was published as part of HackerNoon’s Business Blogging Program. For more details, please click here . here

Brander Group's Vendor-Neutral Strategy Shields Clients from Network Security Vulnerabilities

Η στρατηγική ουδέτερης προμηθευτή του Ομίλου Brander προστατεύει τους πελάτες από ευπάθειες ασφάλειας δικτύου

La estrategia de Brander Group de no tener proveedores protege a los clientes de las vulnerabilidades de seguridad de la red

Brander Group-ի վաճառողի կողմից չեզոք ռազմավարությունը պաշտպանում է հաճախորդներին ցանցային անվտանգության խոցելիությունից

La strategia Vendor-Neutral di Brander Group protegge i clienti dalle vulnerabilità della sicurezza di rete

Brander Group のベンダー中立戦略により、クライアントをネットワーク セキュリティの脆弱性から保護

Brander Group-ის გამყიდველის ნეიტრალური სტრატეგია იცავს კლიენტებს ქსელის უსაფრთხოების დაუცველობისგან

ຍຸດທະສາດຜູ້ຂາຍ-ເປັນກາງຂອງກຸ່ມ Brander ປົກປ້ອງລູກຄ້າຈາກຊ່ອງໂຫວ່ຄວາມປອດໄພຂອງເຄືອຂ່າຍ

Стратегијата за неутрална добавувач на Brander Group ги штити клиентите од ранливости на мрежната безбедност

د برانډر ګروپ د پلورونکي - بې طرفه ستراتیژي پیرودونکي د شبکې امنیت زیانمننې څخه ساتي

Brander Grupo kaqpa Vendor-Neutral Estrategian Clientekunata Llika Seguridad Vulnerabilidades kaqmanta harkan

Брандер Гроуп-ова стратегија неутрална према добављачима штити клијенте од пропуста у мрежној безбедности

ናይ ብራንደር ግሩፕ ሸያጢ-ገለልተኛ ስትራተጂ ንዓማዊል ካብ ተቓላዕነት ድሕነት መርበብ ይከላኸለሎም

In an era when digital technologies shape nearly every aspect of our lives, it is easy to assume these tools are neutral — just a means to an end. Be it smartphones, social media, or the latest blockchain innovation, the overall presumption is that these are all just tools, ready to be used for good or bad according to human intentions. But what if that assumption is wrong? What if these technologies were indeed inherently political, meant to orchestrate our ways of living, working, and interacting, whether we realized it or not? The Myth of Technological Neutrality For many of us, the idea that technology is neutral is a comforting notion. It lets us focus on the exciting benefits of innovations without worrying too much about the potential downsides. However, some scholars argue that this belief is simply a myth. Technologies, they say, are not just passive tools. They are active agents in shaping society and carrying with them values, ideologies, and power dynamics. One of the best arguments against technological neutrality came from the political theorist Langdon Winner. He asked, "Do artifacts have politics?" Work like Winner's has argued that technologies are far from neutral—they can enforce specific social structures and power relations simply by how they are designed and implemented. For instance, blockchains are a digital ledger technology that has swept the planet by storm for their purported potential to decentralize politics by removing the need for central authorities like banks or governments. On the face of it, blockchain technology could serve as a perfect candidate to promote equality and fairness. However, things start getting complicated once we get beyond the surface level. Blockchain: A Tool for Decentralization or a New Means of Control? Blockchain technology, which is best known for powering cryptocurrencies like Bitcoin, has been widely hailed as potentially revolutionary—a tool that will democratize access to financial services and empower individuals. By allowing peers to transact with each other without having to go through an intermediary, blockchain promises to reduce reliance on centralized institutions and give people more control over their assets. Yet, as with any technology, the way blockchain is used and who controls its implementation can lead to very different outcomes. For instance, though it has the potential to disrupt traditional power structures, it can also entrench existing inequalities. If only a tiny tech-savvy elite can understand and control blockchain systems, then the technology may actually serve to further centralize power rather than distribute it. This brings us to an important question: How can we ensure that blockchain and other digital technologies are used to empower people, rather than just serving the interests of those already in power? The Civic Potential of Blockchain: The Case of CommonsHood One intriguing example of blockchain being used for social good is the CommonsHood project in Turin, Italy. Developed by the Computer Science Department at the University of Turin, CommonsHood is a blockchain-based wallet app designed to empower local communities by enabling them to manage resources in a decentralized way. In contrast to many blockchain applications that focus on global financial markets, CommonsHood is based on the specific needs of local communities. The project aims to equip citizens, local businesses, and civic organizations with a tool to issue and manage digital tokens for several uses: fundraising for a community project, trading services among residents, or even rewarding volunteers. Its focus on social equality and inclusion sets CommonsHood apart from other blockchain projects. The project is designed to be accessible to everyone, not just those with advanced technical skills. This design choice ensures that all members of the community, from small business owners to civic organizations, can participate in and benefit from the decentralized economy that CommonsHood seeks to create. For instance, a local civic association could use CommonsHood to crowd-fund a new community center. Instead of using something like Kickstarter, which takes a significant cut of the funds raised on its platform, the association could issue cryptographic tokens through CommonsHood. These could be redeemed by people who want to access services from the new center. By offering an alternative to traditional fundraising and resource management platforms, CommonsHood is hoping to foster a sense of shared ownership and community engagement, ultimately working to build stronger, more resilient local economies. Challenges of Blockchain in Civic Applications However, the potential to use blockchain for social good has its challenges. One of the most important concerns is that not all people possess sufficient digital literacy or access to technology, such as smartphones, or a reliable internet connection to meaningfully engage in a blockchain-based system. Instead, this raises important questions about who those technologies will empower and who might get left behind. Moreover, in addition to the democratic and equity-promoting potential of projects like CommonsHood, they also point to a larger question of how market forces can co-opt digital technologies. Even in well-meaning projects, there still remains a possibility that the technology will simply serve the interests of private interests over the public good. The Politics of Technology Design At the heart of these challenges lies the fact that the design of technology is never neutral. Decisions made in developing a technology, in what it is designed to do, who it is designed for, and how it is is deployed are inherently political. These decisions reflect value judgments and priorities that the technology reinforces, which can have far-reaching implications for society. In the case of blockchain, this means that the way the technology is designed and implemented will determine whether it leads to greater democratization or increased centralization of power. Suppose we want blockchain to live up to its potential as a tool for social good. In that case, we should  look carefully at design choices to ensure that they align with principles of fairness, accessibility, and democratic control. Conclusion: Toward a More Equitable Digital Future With the continued development of blockchain technology, we need to be vigilant not to lose sight of the social and political outcomes of the innovations. Projects like CommonsHood offer a glimpse of how blockchain can empower communities and foster social equity. But for it to become effective, more than just technical innovation has to be done. A commitment must be made toward designing technologies that are inclusive, transparent, and accountable. Ultimately, the future of blockchain and any other digital technology lies in the decisions that are made today. Recognizing the political nature of these technologies and taking steps to ensure they will be put into use for the public good can help shape this future to be more fair and democratic. In an era when digital technologies shape nearly every aspect of our lives, it is easy to assume these tools are neutral — just a means to an end. Be it smartphones, social media, or the latest blockchain innovation, the overall presumption is that these are all just tools, ready to be used for good or bad according to human intentions. But what if that assumption is wrong? What if these technologies were indeed inherently political, meant to orchestrate our ways of living, working, and interacting, whether we realized it or not? The Myth of Technological Neutrality The Myth of Technological Neutrality For many of us, the idea that technology is neutral is a comforting notion. It lets us focus on the exciting benefits of innovations without worrying too much about the potential downsides. However, some scholars argue that this belief is simply a myth. Technologies, they say, are not just passive tools. They are active agents in shaping society and carrying with them values, ideologies, and power dynamics. One of the best arguments against technological neutrality came from the political theorist Langdon Winner. He asked, " Do artifacts have politics? " Work like Winner's has argued that technologies are far from neutral—they can enforce specific social structures and power relations simply by how they are designed and implemented. Do artifacts have politics? For instance, blockchains are a digital ledger technology that has swept the planet by storm for their purported potential to decentralize politics by removing the need for central authorities like banks or governments. On the face of it, blockchain technology could serve as a perfect candidate to promote equality and fairness. However, things start getting complicated once we get beyond the surface level. Blockchain: A Tool for Decentralization or a New Means of Control? Blockchain: A Tool for Decentralization or a New Means of Control? Blockchain technology, which is best known for powering cryptocurrencies like Bitcoin , has been widely hailed as potentially revolutionary—a tool that will democratize access to financial services and empower individuals. By allowing peers to transact with each other without having to go through an intermediary, blockchain promises to reduce reliance on centralized institutions and give people more control over their assets. Bitcoin Yet, as with any technology, the way blockchain is used and who controls its implementation can lead to very different outcomes. For instance, though it has the potential to disrupt traditional power structures, it can also entrench existing inequalities. If only a tiny tech-savvy elite can understand and control blockchain systems, then the technology may actually serve to further centralize power rather than distribute it. This brings us to an important question: How can we ensure that blockchain and other digital technologies are used to empower people, rather than just serving the interests of those already in power? This brings us to an important question: How can we ensure that blockchain and other digital technologies are used to empower people, rather than just serving the interests of those already in power? The Civic Potential of Blockchain: The Case of CommonsHood The Civic Potential of Blockchain: The Case of CommonsHood One intriguing example of blockchain being used for social good is the CommonsHood project in Turin, Italy. Developed by the Computer Science Department at the University of Turin, CommonsHood is a blockchain-based wallet app designed to empower local communities by enabling them to manage resources in a decentralized way. CommonsHood In contrast to many blockchain applications that focus on global financial markets, CommonsHood is based on the specific needs of local communities. The project aims to equip citizens, local businesses, and civic organizations with a tool to issue and manage digital tokens for several uses: fundraising for a community project, trading services among residents, or even rewarding volunteers. Its focus on social equality and inclusion sets CommonsHood apart from other blockchain projects. The project is designed to be accessible to everyone, not just those with advanced technical skills. This design choice ensures that all members of the community, from small business owners to civic organizations, can participate in and benefit from the decentralized economy that CommonsHood seeks to create. For instance, a local civic association could use CommonsHood to crowd-fund a new community center. Instead of using something like Kickstarter, which takes a significant cut of the funds raised on its platform, the association could issue cryptographic tokens through CommonsHood. These could be redeemed by people who want to access services from the new center. By offering an alternative to traditional fundraising and resource management platforms, CommonsHood is hoping to foster a sense of shared ownership and community engagement, ultimately working to build stronger, more resilient local economies. Challenges of Blockchain in Civic Applications Challenges of Blockchain in Civic Applications However, the potential to use blockchain for social good has its challenges. One of the most important concerns is that not all people possess sufficient digital literacy or access to technology, such as smartphones, or a reliable internet connection to meaningfully engage in a blockchain-based system. Instead, this raises important questions about who those technologies will empower and who might get left behind. Moreover, in addition to the democratic and equity-promoting potential of projects like CommonsHood, they also point to a larger question of how market forces can co-opt digital technologies. Even in well-meaning projects, there still remains a possibility that the technology will simply serve the interests of private interests over the public good. The Politics of Technology Design The Politics of Technology Design At the heart of these challenges lies the fact that the design of technology is never neutral. Decisions made in developing a technology, in what it is designed to do, who it is designed for, and how it is is deployed are inherently political. These decisions reflect value judgments and priorities that the technology reinforces, which can have far-reaching implications for society. In the case of blockchain, this means that the way the technology is designed and implemented will determine whether it leads to greater democratization or increased centralization of power. Suppose we want blockchain to live up to its potential as a tool for social good. In that case, we should  look carefully at design choices to ensure that they align with principles of fairness, accessibility, and democratic control. Conclusion: Toward a More Equitable Digital Future Conclusion: Toward a More Equitable Digital Future With the continued development of blockchain technology, we need to be vigilant not to lose sight of the social and political outcomes of the innovations. Projects like CommonsHood offer a glimpse of how blockchain can empower communities and foster social equity. But for it to become effective, more than just technical innovation has to be done. A commitment must be made toward designing technologies that are inclusive, transparent, and accountable. Ultimately, the future of blockchain and any other digital technology lies in the decisions that are made today. Recognizing the political nature of these technologies and taking steps to ensure they will be put into use for the public good can help shape this future to be more fair and democratic.

Stories, News & Mentions about NEU

Neu’s Stories on HackerNoon

Neu's News & Mentions