The Thesis Review [26] Kevin Ellis - Algorithms for Learning to Induce Programs
May 29, 2021
Kevin Ellis, an assistant professor at Cornell and a research scientist at Common Sense Machines, dives into the intriguing world of AI and program synthesis. He discusses his groundbreaking work on DreamCoder, which automates the creation of programming libraries using neural networks. Ellis explores the fusion of AI with natural language and cognitive models, emphasizing Bayesian approaches that mirror human cognition. He shares insights on bridging program synthesis with theorem proving, highlighting the importance of reusable abstractions in machine learning.
AI Snips
Chapters
Books
Transcript
Episode notes
Bayesian view on human cognition
- Bayesian inference can be seen as either a useful engineering tool or a model of brain computation.
- Evolutionary pressures may lead brains to behave "as if" doing Bayesian inference without explicit implementation.
Programs for holistic visual perception
- Programs can explain high-level visual patterns and enable extrapolation beyond raw perception.
- Using known symbol types simplifies parsing but limits symbol discovery and generalization in perception.
Hybrid neural-symbolic synthesis approach
- Use learned neural networks to guide symbolic program synthesizers instead of replacing them.
- Leverage symbolic solvers for what they're good at, and neural networks for heuristic guidance.
