Neural networks for abstraction and reasoning.

Abstract

For half a century, artificial intelligence research has attempted to reproduce the human qualities of abstraction and reasoning - creating computer systems that can learn new concepts from a minimal set of examples, in settings where humans find this easy. While specific neural networks are able to solve an impressive range of problems, broad generalisation to situations outside their training data has proved elusive. In this work, we look at several novel approaches for solving the Abstraction & Reasoning Corpus (ARC). This is a dataset of abstract visual reasoning tasks introduced to test algorithms on broad generalization. Despite three international competitions with $100,000 in prizes, the best algorithms still fail to solve a majority of ARC tasks. The best solvers today rely on complex hand-crafted rules, without using machine learning at all. We revisit whether recent advances in neural networks allow progress on this task, or whether an entirely different class of models are required. First, we adapt the DreamCoder neurosymbolic reasoning solver to ARC. DreamCoder automatically writes programs in a bespoke domain-specific language to perform reasoning, using a neural network to mimic human intuition. We present the Perceptual Abstraction and Reasoning Language (PeARL) language, which allows DreamCoder to solve ARC tasks, and propose a new recognition model that allows us to significantly improve on the previous best implementation. We also propose a new encoding and augmentation scheme that allows large language models (LLMs) to solve ARC tasks, and find that the largest models can solve some ARC tasks. LLMs are able to solve a different group of problems to state-of-the-art solvers, and provide an interesting way to complement other approaches. We perform an ensemble analysis, combining systems to achieve better results than any system alone and analysing individual strengths. However, it is sobering to see that approaches based on neural networks still lag behind existing hand-crafted solvers, and we suggest avenues for future improvements. Our findings with the ensemble model may indicate that a diversity of methods might be necessary to solve problems in ARC. Humans likely employ diverse strategies to solve ARC. Studies involving human participants to identify the strategies they employ to solve ARC could provide valuable insights for future AI approaches. Finally, we publish the arckit Python library to make future research on ARC easier.