The following is an abstract of some thoughts I had from the RWKV podcast found here: https://www.latent.space/p/rwkv#details Why do Alternatives Matter? Transformer architectures are currently at their peak with the AI revolution of 2023. However, in the rush to adopt them due to their recent success, it's easy to ignore alternatives that we can learn from. As engineers, we should not take a one-size-fits-all approach, using the same hammer for every nail. We should find trade-offs in every solution. Failure to do so would trap us within the limitations of one particular platform while remaining blissfully unaware of all other alternatives. This could potentially hold back development for years. This problem is not unique to . It's a historical pattern that has repeated itself since ancient times up to the modern day. AI A page from the SQL war history... A notable recent example in software development was the trend towards NoSQL when SQL servers started hitting physical limitations. Startups everywhere made for "scale" reasons despite being nowhere near those scales. the move to NoSQL However, over time, as eventual consistency and NoSQL management overhead emerged, and hardware capabilities made giant leaps in SSD speed and capacity, we've seen a recent trend back to SQL servers for their simplicity and now sufficient scalability for 90%+ of startups. Does this mean SQL is better than NoSQL or vice versa? No, it simply means each technology has preferred use cases with pros/cons and learning points that can cross-pollinate among similar technologies. I hope you view this article as not an attack on transformers, but a healthy appreciation on alternatives. With that, lets get on to the main topic What is the biggest pain point for existing transformer architecture? Generally, this covers compute, context length, datasets, and alignment. For this discussion, we'll focus on compute and context length: Quadratic computational costs due to O(N^2) increase per token used/generated. This makes context sizes >100k incredibly expensive, which impacts inference and training. The current exacerbates this issue. GPU shortage Context size limits attention, heavily restricting "smart agent" use cases (like ), and forcing workarounds. Larger contexts need fewer workarounds. smol-dev So, how do we solve this? Introducing RWKV - a linear transformer/modern large-scale RNN The following summarises the RWKV paper found: here For a deeper technical dive, see the: wiki here , along with , are among the first of a new class called RWKV Microsoft RetNet "Linear Transformers." It tackles the above three limitations directly by supporting the following: Linear computational cost regardless of context size. Lower requirements allow reasonable token/sec output on CPU (especially ARM) in RNN mode. No hard context size limit as an RNN. Any limits in docs are guidelines - you can finetune beyond them. As we continue to scale up AI models to large context sizes. From 100k and beyond. The quadratic computational costs start to increase exponentially. Linear Transformers, however, instead of abandoning the Recurrent Neural Network architecture and their bottlenecks, which forced them to be replaced by transformers. Took lessons learned from making transformers scalable, re-engineered RNNs to work similarly to transformers, and removed all those bottlenecks. Bringing them back into the race with transformers in terms of training speeds. Allowing them to both operate efficiently at O(N) cost while scaling past a billion parameters in training. All while maintaining similar performance levels. As you scale quadratically vs. linear, you will hit over 10x at a 2k token count & over 100x at a 100k token length. At 14B parameters, RWKV is the largest open-source linear transformer, on par with GPT-NeoX and others on similar datasets (e.g. The Pile). But what does this mean, in simpler terms? Pros 10x++ cheaper inference/training than transformers at large context sizes Can run slowly on very limited hardware in RNN mode Similar performance to transformers on the same datasets No technical context size limit as RNN (infinite context!) Cons Sliding window problem, with lossy memory past a certain point Not proven to scale beyond 14B parameters yet Less optimized and adopted than transformers So, while RWKV isn't yet at 60B+ parameter scales like LLaMA2, with the right support/resources, it has the potential to get there at lower cost and larger contexts, especially as expert models trend towards smaller and more efficient. Consider it if efficiency matters for your use case. But it's not the final solution - healthy alternatives are still key. Also one last thing i forget to mentioned: Linear Transformers typically do not use classic Transformer attention, but an approximate, meaning ... Attention is Not All You Need Other alternatives and their benefits we should probably learn from : Slow to train for text but extremely resilient to multi-epoch training. Figuring out why could help mitigate Diffusion Models the token crisis : techniques could be used to train desired skill sets to a specific goal without datasets. Even for text-based models Generative Adversarial Networks / Agents I am a code contributor to the RWKV project and an active member of the community. However, it does not mean I view RWKV as the final answer; even if in the future, where "Linear Transformers" beat out traditional "Transformer" networks in the future, research into alternatives is still important for healthy ecosystem development. Disclosure: I strongly believe there is no such thing as "one architecture to rule them all.” Additionally, I am aware there have since been developments like TransformerXL to more efficiently scale transformers, somewhere between quadratic cost and linear cost. However, with very limited known implementation of this method in public, it is hard for me to evaluate its pros and cons. Also published . here

The best podcasts on the Internet archived and shared on HackerNoon.

Understanding Vector Embeddings for AI-Powered Search

Sleep Well, with Low-Code UI test automation @ uilicious.com

Introducing RWKV: The Rise of Linear Transformers and Exploring Alternatives

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

How AI/ChatGPT Dreams in 2022

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Noonification: The Metaverse is a Sh*tshow (11/2/2022)

100 Days of AI Day 1: From Newsletter to Podcast, Leveraging AI for Audio Transformation

How AI/ChatGPT Dreams in 2022

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

The Noonification: White Man (11/26/2022)

The Noonification: The Metaverse is a Sh*tshow (11/2/2022)

100 Days of AI Day 1: From Newsletter to Podcast, Leveraging AI for Audio Transformation

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps