The Emergent Symbolic Structure of Artificial Neural Networks
Computation and LanguageArtificial Intelligence
Summary
The authors explore why neural networks, which use continuous vectors, can perform well on tasks usually thought to require symbolic reasoning, like language and logic. They suggest that neural networks might internally represent information in ways that are similar to symbols, even if it doesn't look like it. They show that the networks' vector representations can be closely matched by symbolic structures without changing the networks' behavior much. This works for both small networks handling lists and large language models across tasks like math, logic, coding, and language. Their findings offer a way to connect traditional symbolic views of intelligence with modern AI.
Neural networksSymbolic representationsLanguage modelsVector embeddingsLogical reasoningArithmeticComputer codeInternal representationsSymbolic structureArtificial intelligence
Authors
R. Thomas McCoy, Paul Soulos, Tal Linzen, Paul Smolensky
Abstract
Modern systems in artificial intelligence (AI) somehow excel in domains for which they seem poorly suited. Intelligence has traditionally been modeled as operating over structured combinations of symbols, such as logical formulas. However, the strongest modern AI systems are based on neural networks, which instead represent information in continuous vectors. Vectors seem inadequate for capturing the structure of language, logic, and other cognitive domains, yet neural networks achieve impressive performance in these areas. How do they do it? In this work, we propose a potential answer: Despite appearances, perhaps the internal representations of neural networks implicitly realize symbolic structure. In support of this hypothesis, we show that the vector representations of a variety of neural networks can be closely approximated with symbolic structures: we can replace the network's entire representation-generating process with a closed-form equation instantiating a symbolic structure, and the network's behavior remains largely unchanged. This finding holds for both small-scale neural networks trained to manipulate lists as well as large language models (LLMs) operating in four domains that are central in symbolic traditions: arithmetic, logic, computer code, and language. Further, our symbolic approximation allows us to modify an LLM's behavior in targeted ways via precise interventions on its internal representations, showing that the LLM's behavior is reliant on the symbolic structures we have identified. This work provides a potential way to reconcile longstanding symbolic conceptions of intelligence with the vector-based nature of modern AI.