Polysemy of Synthetic Neurons Towards a New Type of Explanatory Categorical Vector Spaces

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Pichat, Michael, Pogrund, William, Pichat, Paloma, Poumay, Judicael, Gasparian, Armanouche, Demarchi, Samuel, Corbet, Martin, Georgeon, Alois, Veillet-Guillem, Michael
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913833912631296
author Pichat, Michael
Pogrund, William
Pichat, Paloma
Poumay, Judicael
Gasparian, Armanouche
Demarchi, Samuel
Corbet, Martin
Georgeon, Alois
Veillet-Guillem, Michael
author_facet Pichat, Michael
Pogrund, William
Pichat, Paloma
Poumay, Judicael
Gasparian, Armanouche
Demarchi, Samuel
Corbet, Martin
Georgeon, Alois
Veillet-Guillem, Michael
contents The polysemantic nature of synthetic neurons in artificial intelligence language models is currently understood as the result of a necessary superposition of distributed features within the latent space. We propose an alternative approach, geometrically defining a neuron in layer n as a categorical vector space with a non-orthogonal basis, composed of categorical sub-dimensions extracted from preceding neurons in layer n-1. This categorical vector space is structured by the activation space of each neuron and enables, via an intra-neuronal attention process, the identification and utilization of a critical categorical zone for the efficiency of the language model - more homogeneous and located at the intersection of these different categorical sub-dimensions.
format Preprint
id arxiv_https___arxiv_org_abs_2505_07831
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Polysemy of Synthetic Neurons Towards a New Type of Explanatory Categorical Vector Spaces
Pichat, Michael
Pogrund, William
Pichat, Paloma
Poumay, Judicael
Gasparian, Armanouche
Demarchi, Samuel
Corbet, Martin
Georgeon, Alois
Veillet-Guillem, Michael
Computation and Language
Artificial Intelligence
Machine Learning
The polysemantic nature of synthetic neurons in artificial intelligence language models is currently understood as the result of a necessary superposition of distributed features within the latent space. We propose an alternative approach, geometrically defining a neuron in layer n as a categorical vector space with a non-orthogonal basis, composed of categorical sub-dimensions extracted from preceding neurons in layer n-1. This categorical vector space is structured by the activation space of each neuron and enables, via an intra-neuronal attention process, the identification and utilization of a critical categorical zone for the efficiency of the language model - more homogeneous and located at the intersection of these different categorical sub-dimensions.
title Polysemy of Synthetic Neurons Towards a New Type of Explanatory Categorical Vector Spaces
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2505.07831