Neuro-symbolic program synthesis and applications to human-level concept learning
We present a novel program synthesis technique which learns to infer programs from examples using self-supervision and reinforcement learning. Our model (https://arxiv.org/pdf/1906.04604.pdf) formulates program synthesis as an iterative search. Given a specification, we predict a small code segment with a learned policy, symbolically execute the code segment to produce a partial execution state, and assess this partial state using a learned value function. We test our model on string editing tasks and simple 2D&3D geometric models.
In the second half of the talk, we will discuss applications of this program synthesis method to human concept learning tasks. We will show preliminary results applying our system to an artificial language learning task from Lake et al (https://cims.nyu.edu/~brenden/papers/LakeEtAl2019CogSci.pdf). We show that our system, like humans, is able to infer the rules for this novel language from only 14 examples.
Beyond accuracy and efficiency: How our goals shape learning and exploration
Rational theories of learning typically assume that agents' motivations are purely epistemic (i.e. to gain information or reduce uncertainty). However, sometimes this leads to intractable search problems in hypothesis generation, search, and evaluation. Might considering additional motivations provide us with useful constraints that support effective learning? Here, we consider how goals other than obtaining accurate information and performing efficient actions can account for behaviors in 1) judgment and evaluation and 2) search and exploration.
In Study 1 (N=244) we find that 4-7-year-old children override documented preferences for accurate, verified, and confident claims to endorse unverified conjectures, but only if the conjectures answer an unresolved question. In Study 2 (N=110) we find that children aged 4 and 5 years, but not 3-year-olds, use the size of search space to make efficient plans. However, children spontaneously choose inefficient actions when given the chance to play. These results suggest that even young children make trade-offs among various criteria for hypothesis evaluation and selection, and are sensitive to context-dependent motivations for determining which criteria to prioritize and which variables to treat as costs versus reward. We end by presenting some directions for future research.