Shift AI 2020: Using AI for automatic synthesis | Boris Cergol (Comtrade Digital Services)
0 likes84 views
The document discusses the application of AI and machine learning in the automatic synthesis of computer programs, covering topics like program specification, neural program synthesis, and reinforcement learning. It highlights challenges in program synthesis, such as specification interpretation and program space intractability, while presenting various methods and hybrid learning approaches for program generation. The document also touches on commercial solutions and future directions in AI program synthesis, including the synthesis of programs from multimodal inputs.
![Example of neural-symbolic hybrid
CounterExample Guided Neural Synthesis [Polgreen 2020]
1) Use symbolic SMT solver to turn logical specification into I/O examples.
2) Use neural network to generate candidate programs.
3) Use symbolic SMT solver to verify candidate programs.
[Polgreen 2020] Polgreen, Elizabeth, Ralph Abboud, and Daniel Kroening. "CounterExample Guided Neural
Synthesis." arXiv preprint arXiv:2001.09245 (2020).](https://image.slidesharecdn.com/shiftai-boriscergol-200427143156/85/Shift-AI-2020-Using-AI-for-automatic-synthesis-Boris-Cergol-Comtrade-Digital-Services-8-320.jpg)
![Example of neural-symbolic hybrid
Learning to Infer Program Sketches [Nye 2019]
1) When programming, humans seem to combine learned patterns and explicit reasoning.
2) Program sketch: schematic outline of full program
3) Neural network generates candidate sketches.
4) Sketches are converted to full programs using symbolic synthesis.
[Nye 2019] Nye, Maxwell, et al. "Learning to infer program sketches." arXiv preprint
arXiv:1902.06349 (2019).](https://image.slidesharecdn.com/shiftai-boriscergol-200427143156/85/Shift-AI-2020-Using-AI-for-automatic-synthesis-Boris-Cergol-Comtrade-Digital-Services-9-320.jpg)
![Reinforcement learning
Write, Execute, Assess: Program Synthesis with a REPL [Ellis 2019]
1) Use reinforcement learning to explore semantic space of programs.
2) Inspired by human trial-and-error programming and AlphaGo architecture.
3) Policy proposes code to write, value function evaluates long-term prospects of code.
4) Interpreter bridges gap between syntax and semantics.
[Ellis 2019] Ellis, Kevin, et al. "Write, execute, assess: Program synthesis with a repl." Advances in Neural
Information Processing Systems. 2019.](https://image.slidesharecdn.com/shiftai-boriscergol-200427143156/85/Shift-AI-2020-Using-AI-for-automatic-synthesis-Boris-Cergol-Comtrade-Digital-Services-10-320.jpg)
Shift AI 2020: Using AI for automatic synthesis | Boris Cergol (Comtrade Digital Services)
- 1. UsingAI for automatic synthesis of computer programs Boris Cergol, PhD Head of AI, Comtrade Digital Services Shift AI
- 2. Introduction Machine learning Software engineering CI/CD, versioning, QA, agile, data management, microservices, containters,… AutoML, NAS ?
- 3. Automatic program generation Program specification Program specification Human-readable program satisfying specification Non-interpretable program satisfying specification Program synthesis Program induction
- 4. Program synthesis with different specifications Specification type Program synthesis type Input-output (I/O) pairs Programming-by-example (PBE) Incomplete code, complete Code completion Faulty code, repair Program repair Code, target language Code migration Natural language Code generation/semantic parsing Demonstration Programming-by-demonstration (PBD) Code, natural language, screenshots, speech,… Multimodal program synthesis
- 5. Why is program synthesis hard? Program space intractability Domain specific languages Lack of data Synthetic data Generalization Specification multi-modality User intent interpretation Program evaluation metrics Discrete representatio ns Long dependencies Large vocabularies Rich structure Sensitivity to changes
- 6. Neural program synthesis Deep learning often applied to program synthesis through encoder-decoder seq2seq type models. I/O pairs Encoder Embedding Decoder Program
- 7. Neural vs Symbolic Source: https://github.com/crazydonkey200/neural-symbolic-machines Marcus, Gary. "The next decade in AI: four steps towards robust artificial intelligence." arXiv preprint arXiv:2002.06177 (2020).
- 8. Example of neural-symbolic hybrid CounterExample Guided Neural Synthesis [Polgreen 2020] 1) Use symbolic SMT solver to turn logical specification into I/O examples. 2) Use neural network to generate candidate programs. 3) Use symbolic SMT solver to verify candidate programs. [Polgreen 2020] Polgreen, Elizabeth, Ralph Abboud, and Daniel Kroening. "CounterExample Guided Neural Synthesis." arXiv preprint arXiv:2001.09245 (2020).
- 9. Example of neural-symbolic hybrid Learning to Infer Program Sketches [Nye 2019] 1) When programming, humans seem to combine learned patterns and explicit reasoning. 2) Program sketch: schematic outline of full program 3) Neural network generates candidate sketches. 4) Sketches are converted to full programs using symbolic synthesis. [Nye 2019] Nye, Maxwell, et al. "Learning to infer program sketches." arXiv preprint arXiv:1902.06349 (2019).
- 10. Reinforcement learning Write, Execute, Assess: Program Synthesis with a REPL [Ellis 2019] 1) Use reinforcement learning to explore semantic space of programs. 2) Inspired by human trial-and-error programming and AlphaGo architecture. 3) Policy proposes code to write, value function evaluates long-term prospects of code. 4) Interpreter bridges gap between syntax and semantics. [Ellis 2019] Ellis, Kevin, et al. "Write, execute, assess: Program synthesis with a repl." Advances in Neural Information Processing Systems. 2019.
- 11. Some more reinforcement 1) Xu, Yifan, et al. "Neural Program Synthesis By Self-Learning." arXiv preprint arXiv:1910.05865 (2019) Learn to synthesize assembly code without any human knowledge. 2) Pierrot, Thomas, et al. "Learning compositional neural programs with recursive tree search and planning." Advances in Neural Information Processing Systems. 2019. Learn to synthesize a hierarchic library of programs required to solve a problem. 3) Liu, Xiao, et al. "Reinforcement Compiler Fuzzing." (2019) Learn an agent that can generate programs to perform fuzz testing of compilers.
- 12. Commercial solutions? Flash Fill in Excel Prose IntelliCode in VS Aroma Big Tech CodeGuru Startups ?
- 13. TF-Coder by Google Programming-by-example system that automatically synthesizes tensor manipulation programs from input/output examples and natural language descriptions. “TF-Coder often produces solutions that are simpler and more elegant than those written by TensorFlow experts (including the authors of this paper).” TF-Coder: Program Synthesis for Tensor Manipulations
- 14. Moving beyond software engineering Duan, Xuguang, et al. "Watch, Reason and Code: Learning to Represent Videos Using Program." Proceedings of the 27th ACM International Conference on Multimedia. 2019. Synthesis of programs that describe processes carried out in videos
- 15. Moving beyond software engineering Chollet, Franccois. “On the Measure of Intelligence.” ArXiv abs/1911.01547 (2019) “The measure of the success of our message will be its ability to divert the attention of some part of the community interested in general AI, away from surpassing humans at tests of skill, towards investigating the development of human-like broad cognitive abilities, through the lens of program synthesis, Core Knowledge priors, curriculum optimization, information efficiency, and achieving extreme generalization through strong abstraction”
- 16. Thank you for you attention Boris Cergol boris.cergol@comtrade.com https://comtradedigital.com/ https://www.comtrade.ai/