Top-Down Synthesis for Library Learning

Bowers, Matthew T.; Olausson, Theo X.; Wong, Catherine; Grand, Gabriel; Tenenbaum, Joshua B.; Ellis, Kevin; Solar-Lezama, Armando

doi:10.1145/3571234

Cited by 15 publications

(4 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These methods relax the context-free assumption used in traditional Bayesian symbolic models, and jointly infers both the posterior over concept 'programs', and a latent library that defines this posterior. This notion of concept libraries is attracting increasing attention across cognitive science and generative AI (Bowers et al 2023;Tian et al 2020;Wang et al 2023;Wong et al 2022).…”

Section: Resource-rational Library Learningmentioning

confidence: 99%

Constructing Deep Concepts through Shallow Search

Zhao,

Lucas,

Bramley

2024

AAAI-SS

View full text Add to dashboard Cite

We propose bootstrap learning as a computational account for why human learning is modular and incremental, and identify key components of bootstrap learning that allow artificial systems to learn more like people. Originated from developmental psychology, bootstrap learning refers to people's ability to extend and repurpose existing knowledge to create new and more powerful ideas. We view bootstrap learning as a solution of how cognitively-bounded reasoners grasp complex environmental dynamics that are far beyond their initial capacity, by searching ‘locally’ and recursively to extend their existing knowledge. Drawing from techniques of Bayesian library learning and resource rational analysis, we propose a computational modeling framework that achieves human-like bootstrap learning performance in inductive conceptual inference. In addition, we demonstrate modeling and behavioral evidence that highlights the double-edged sword of bootstrap learning, such that people processing the same information in different batch orders could induce drastically different causal conclusions and generalizations, as a result of the different sub-concepts they construct in earlier stages of learning.

show abstract

Section: Resource-rational Library Learningmentioning

confidence: 99%

Constructing Deep Concepts through Shallow Search

Zhao,

Lucas,

Bramley

2024

AAAI-SS

View full text Add to dashboard Cite

show abstract

“…As the number of possible refactorings grows combinatorially with program size, we needed a new data structure for representing and manipulating sets of refactorings, which we designed by combining ideas from version space algebras [20][21][22] and equivalence graphs [23] (described in our companion manuscript [2]). Recent work improved upon our original refactoring algorithm by making it more expressive [24] as well as orders of magnitude faster [25].…”

Section: Wake/sleep Program Learningmentioning

confidence: 99%

DreamCoder: growing generalizable, interpretable knowledge with wake–sleep Bayesian program learning

Ellis

Wong

Nye

et al. 2023

Phil. Trans. R. Soc. A.

Self Cite

View full text Add to dashboard Cite

Expert problem-solving is driven by powerful languages for thinking about problems and their solutions. Acquiring expertise means learning these languages—systems of concepts, alongside the skills to use them. We present DreamCoder, a system that learns to solve problems by writing programs. It builds expertise by creating domain-specific programming languages for expressing domain concepts, together with neural networks to guide the search for programs within these languages. A ‘wake–sleep’ learning algorithm alternately extends the language with new symbolic abstractions and trains the neural network on imagined and replayed problems. DreamCoder solves both classic inductive programming tasks and creative tasks such as drawing pictures and building scenes. It rediscovers the basics of modern functional programming, vector algebra and classical physics, including Newton’s and Coulomb’s laws. Concepts are built compositionally from those learned earlier, yielding multilayered symbolic representations that are interpretable and transferrable to new tasks, while still growing scalably and flexibly with experience. This article is part of a discussion meeting issue ‘Cognitive artificial intelligence’.

show abstract

“…While the GNN of a G-SSNN can be trained with standard gradient-based optimization techniques, its graph structure is a hyperparameter whose values must be explored with search. To conduct this search, I repurpose the cycle of library learning and distributional program search pioneered by the DreamCoder system [Ellis et al, 2021] within an evolutionary framework, incorporating the improved STITCH library learning tool of Bowers et al [2023] and heap search algorithm of Matricon et al [2022]. As in DreamCoder, I perform rounds of library learning and distributional program search, after which I evaluate the current crop of programs on the task.…”

Section: Evolutionary Frameworkmentioning

confidence: 99%

Symbolic Synthesis of Neural Networks

Whitehouse¹

2023

Preprint

View full text Add to dashboard Cite

Neural networks adapt very well to distributed and continuous representations, but struggle to generalize from small amounts of data. Symbolic systems commonly achieve data efficient generalization by exploiting modularity to benefit from local and discrete features of a representation. These features allow symbolic programs to be improved one module at a time and to experience combinatorial growth in the values they can successfully process. However, it is difficult to design a component that can be used to form symbolic abstractions and which is adequately overparametrized to learn arbitrary high-dimensional transformations. I present Graph-based Symbolically Synthesized Neural Networks (G-SSNNs), a class of neural modules that operate on representations modified with synthesized symbolic programs to include a fixed set of local and discrete features. I demonstrate that the choice of injected features within a G-SSNN module modulates the data efficiency and generalization of baseline neural models, creating predictable patterns of both heightened and curtailed generalization. By training G-SSNNs, we also derive information about desirable semantics of symbolic programs without manual engineering. This information is compact and amenable to abstraction, but can also be flexibly recontextualized for other high-dimensional settings. In future work, I will investigate data efficient generalization and the transferability of learned symbolic representations in more complex G-SSNN designs based on more complex classes of symbolic programs. Experimental code and data are available here.

show abstract

Top-Down Synthesis for Library Learning

Cited by 15 publications

References 32 publications

Constructing Deep Concepts through Shallow Search

Constructing Deep Concepts through Shallow Search

DreamCoder: growing generalizable, interpretable knowledge with wake–sleep Bayesian program learning

Symbolic Synthesis of Neural Networks

Contact Info

Product

Resources

About