Table of Links Abstract and 1. Introduction Abstract and 1. Introduction Background & Related Work


Method
3.1 Sampling Small Mutations
3.2 Policy
3.3 Value Network & Search
3.4 Architecture


Experiments
4.1 Environments
4.2 Baselines
4.3 Ablations


Conclusion, Acknowledgments and Disclosure of Funding, and References Background & Related Work Background & Related Work Background & Related Work Method
3.1 Sampling Small Mutations
3.2 Policy
3.3 Value Network & Search
3.4 Architecture Method Method 3.1 Sampling Small Mutations 3.1 Sampling Small Mutations 3.2 Policy 3.2 Policy 3.3 Value Network & Search 3.3 Value Network & Search 3.4 Architecture 3.4 Architecture Experiments
4.1 Environments
4.2 Baselines
4.3 Ablations Experiments 4.1 Environments 4.1 Environments 4.2 Baselines 4.2 Baselines 4.3 Ablations 4.3 Ablations Conclusion, Acknowledgments and Disclosure of Funding, and References Conclusion, Acknowledgments and Disclosure of Funding, and References Conclusion, Acknowledgments and Disclosure of Funding, and References Appendix Appendix A. Mutation Algorithm A. Mutation Algorithm B. Context-Free Grammars B. Context-Free Grammars C. Sketch Simulation C. Sketch Simulation D. Complexity Filtering D. Complexity Filtering E. Tree Path Algorithm E. Tree Path Algorithm F. Implementation Details F. Implementation Details 2 Background & Related Work Neural program synthesis Neural program synthesis is a prominent area of research, in which neural networks generate programs from input-output examples. Early work, such as Parisotto et al. [23], demonstrated the feasibility of this approach. While modern language models can be directly applied to program synthesis, combining neural networks with search strategies often yields better results and guarantees. In this paradigm, the neural network guides the search process by providing proposal distributions or scoring candidate programs. Examples of such hybrid methods include Balog et al. [2], Ellis et al. [12], and Devlin et al. [9]. A key difference from our work is that these methods construct programs incrementally, exploring a vast space of partial programs. Our approach, in contrast, focuses on editing programs, allowing us to both grow programs from scratch and make corrections based on the program execution. Neural program synthesis Neural diffusion Neural diffusion models, a class of generative models, have demonstrated impressive results for modeling high-dimensional data, such as images [14, 22, 31]. A neural diffusion model takes samples from the data distribution (e.g. real-world images), incrementally corrupts the data by adding noise, and trains a neural network to incrementally remove the noise. To generate new samples, we can start with random noise and iteratively apply the neural network to denoise the input. Neural diffusion Diffusion for discrete data Recent work extends diffusion to discrete and structured data like graphs [35], with applications in areas such as molecule design [15, 27, 8]. Notably, Lou et al. [20] proposed a discrete diffusion model using a novel score-matching objective for language modeling. Another promising line of work for generative modeling on structured data is generative flow networks (GFlowNets) [3], where neural models construct structured data one atom at a time. Diffusion for discrete data Diffusion for code generation Singh et al. [30] use a diffusion model for code generation. However, their approach is to first embed text into a continuous latent space, train a continuous diffusion model on that space, and then unembed at the end. This means that intermediate stages of the latent representation are not trained to correspond to actual code. The embedding tokens latch to the nearest embeddings during the last few steps. Diffusion for code generation Direct code editing using neural models has also been explored. Chakraborty et al. [6] use a graph neural network for code editing, trained on a dataset of real-world code patches. Similarly, Zhang et al. [42] train a language model to edit code by modifying or inserting [MASK] tokens or deleting existing tokens. They further fine-tune their model on real-world comments and patches. Unlike these methods, our approach avoids the need for extensive code edit datasets and inherently guarantees syntactic validity through our pretraining task. Program synthesis for inverse graphics We are inspired by previous work by Sharma et al. [28], Ellis et al. [10, 11], which also uses the CSG2D language. Sharma et al. [28] propose a convolutional encoder and a recurrent model to go from images to programs. Ellis et al. [11] propose a method to provide a neural model with the intermediate program execution output in a read–eval– print loop (REPL). Unlike our method, the ability to execute partial graphics programs is a key requirement for their work. Our system operates on complete programs and does not require a custom partial compiler. As mentioned in their work, their policies are also brittle. Once the policy proposes an object, it cannot undo that proposal. Hence, these systems require a large number of particles in a Sequential Monte-Carlo (SMC) sampler to make the system less brittle to mistakes. Program synthesis for inverse graphics Authors:
(1) Shreyas Kapur, University of California, Berkeley (srkp@cs.berkeley.edu);
(2) Erik Jenner, University of California, Berkeley (jenner@cs.berkeley.edu);
(3) Stuart Russell, University of California, Berkeley (russell@cs.berkeley.edu). Authors: Authors (1) Shreyas Kapur, University of California, Berkeley (srkp@cs.berkeley.edu); (2) Erik Jenner, University of California, Berkeley (jenner@cs.berkeley.edu); (3) Stuart Russell, University of California, Berkeley (russell@cs.berkeley.edu). This paper is available on arxiv under CC BY-SA 4.0 DEED license. This paper is available on arxiv under CC BY-SA 4.0 DEED license. available on arxiv available on arxiv

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

Carlo

The Evolution of Code Generation: A Look at Neural Diffusion, GFlowNets, and Direct Editing

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Advances in Photocatalytic Materials through SALSA

Complexity NIMBYs and the Illusion of Transcending Software Complexity Bounds

Improving Chatbot With Code Generation: Building a Context-Aware Chatbot for Publications

Is AI Code Generation An Age of Industrial Revolution for Software Enterprise?

The Easy Way to Begin Test-Driven Development

The Modern AI Stack to Increase Developer Productivity

Advances in Photocatalytic Materials through SALSA

Complexity NIMBYs and the Illusion of Transcending Software Complexity Bounds

Improving Chatbot With Code Generation: Building a Context-Aware Chatbot for Publications

Is AI Code Generation An Age of Industrial Revolution for Software Enterprise?

The Easy Way to Begin Test-Driven Development

The Modern AI Stack to Increase Developer Productivity

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps