GPT-2 Through the Lens of Vector Symbolic Architectures

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Knittel, Johannes, Gangavarapu, Tushaar, Strobelt, Hendrik, Pfister, Hanspeter
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912151925424128
author Knittel, Johannes
Gangavarapu, Tushaar
Strobelt, Hendrik
Pfister, Hanspeter
author_facet Knittel, Johannes
Gangavarapu, Tushaar
Strobelt, Hendrik
Pfister, Hanspeter
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2412_07947
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle GPT-2 Through the Lens of Vector Symbolic Architectures
Knittel, Johannes
Gangavarapu, Tushaar
Strobelt, Hendrik
Pfister, Hanspeter
Machine Learning
Artificial Intelligence
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.
title GPT-2 Through the Lens of Vector Symbolic Architectures
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2412.07947