GPT-2 Through the Lens of Vector Symbolic Architectures

TL;DR

This paper investigates how GPT-2's transformer architecture aligns with vector symbolic architectures, revealing that it employs nearly orthogonal vector operations similar to VSA, which helps explain its neural weights and computational mechanisms.

Contribution

The paper demonstrates a strong resemblance between GPT-2's mechanisms and vector symbolic architectures, providing new insights into its internal operations.

Findings

GPT-2 uses nearly orthogonal vector bundling and binding operations.

VSA principles explain a significant portion of GPT-2's neural weights.

Transformer mechanisms align with vector symbolic architecture concepts.

Abstract

Understanding the general priniciples behind transformer models remains a complex endeavor. Experiments with probing and disentangling features using sparse autoencoders (SAE) suggest that these models might manage linear features embedded as directions in the residual stream. This paper explores the resemblance between decoder-only transformer architecture and vector symbolic architectures (VSA) and presents experiments indicating that GPT-2 uses mechanisms involving nearly orthogonal vector bundling and binding operations similar to VSA for computation and communication between layers. It further shows that these principles help explain a significant portion of the actual neural weights.