Neurosymbolic Program Architecture Search - Methods and Exercises

Explore neurosymbolic program architecture search techniques in this 43-minute tutorial presented by Yisong Yue, Swarat Chaudhuri, and Jennifer Sun. Dive into two primary methods for automatically finding optimal program architectures within domain-specific languages: basic enumerative search and RobustFill. Engage in hands-on exercises to gain practical experience with these methods. Learn about functional idioms, neurosymbolic functional DSLs, and their applications in recurrent neural encoders and counting problems. Understand the benefits of compositionality in top-down program synthesis and explore graph search depictions. Discover key ideas such as estimating "cost to go" and implementing admissible heuristics. Examine the functional representational power assumption and its implications. Conclude with a brief survey of additional methods and tackle a neurosymbolic behavior challenge to reinforce your understanding of the concepts presented.

Outline of Tutorial
Session 2: Neurosymbolic Programming Algorithms
Learning Strategy
Intro to Functional Idioms
A Neurosymbolic Functional DSL
Example: Recurrent Neural Encoder
Example: Counting the number of 5s
Benefit: Compositionality
Top-Down Program Synthesis
Story so Far: Top-Down Synthesis
First Option: Enumerative Search
Graph Search Depiction
Key Idea: Estimate "Cost to Go"
Implementing Admissible Heuristic
Motivating Observation/Assumption: Functional Representational Power
Summary
Code Structure
Neurosymbolic Behavior Challenge