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

2025/04/30 by Pichat, Michael, Pogrund, William, Pichat, Paloma +6
#Artificial Intelligence (cs.AI) #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG)

paper · doi:10.48550/arxiv.2505.07831

Abstract

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.

Citations

• A PennyLane-Centric Dataset to Enhance LLM-based Quantum Code Generation using RAG
• Synthetic Categorical Restructuring large Or How AIs Gradually Extract Efficient Regularities from Their Experience of the World
• The Process of Categorical Clipping at the Core of the Genesis of Concepts in Synthetic Neural Cognition
• How Do Artificial Intelligences Think? The Three Mathematico-Cognitive Factors of Categorical Segmentation Operated by Synthetic Neurons
• The LLM Language Network: A Neuroscientific Approach for Identifying Causally Task-Relevant Units
• Converging to a Lingua Franca: Evolution of Linguistic Regions and Semantics Alignment in Multilingual Large Language Models
• Neuropsychology of AI: Relationship Between Activation Proximity and Categorical Proximity Within Neural Categories of Synthetic Cognition
• Householder Pseudo-Rotation: A Novel Approach to Activation Editing in LLMs with Direction-Magnitude Perspective
• On the Complexity of Neural Computation in Superposition
• Mathematical Models of Computation in Superposition
• The Llama 3 Herd of Models
• Psychology of Artificial Intelligence: Epistemological Markers of the Cognitive Analysis of Neural Networks
• Electrophysiological correlation between executive vigilance and attention network based on cognitive resource control theory
• Sharing Matters: Analysing Neurons Across Languages and Tasks in LLMs
• LLM-assisted Concept Discovery: Automatically Identifying and Explaining Neuron Functions
• Scaling and evaluating sparse autoencoders
• Dissecting the Interplay of Attention Paths in a Statistical Mechanics Theory of Transformers
• A hierarchical Bayesian model for syntactic priming
• Visual Attention in Crisis
• From Cognition to Computation: A Comparative Review of Human Attention and Transformer Architectures
• Let's Focus on Neuron: Neuron-Level Supervised Fine-tuning for Large Language Model
• Massive Activations in Large Language Models
• BEARS Make Neuro-Symbolic Models Aware of their Reasoning Shortcuts
• Transformer Mechanisms Mimic Frontostriatal Gating Operations When Trained on Human Working Memory Tasks
• Can A Cognitive Architecture Fundamentally Enhance LLMs? Or Vice Versa?
• Orthogonal Subspace Learning for Language Model Continual Learning
• Do Compressed LLMs Forget Knowledge? An Experimental Study with Practical Implications
• Llama 2: Open Foundation and Fine-Tuned Chat Models
• Attention-Enhanced Co-Interactive Fusion Network (AECIF-Net) for Automated Structural Condition Assessment in Visual Inspection
• Can LLMs facilitate interpretation of pre-trained language models?
• Finding Neurons in a Haystack: Case Studies with Sparse Probing
• Disentangling Neuron Representations with Concept Vectors
• Evaluating Neuron Interpretation Methods of NLP Models
• Neural Network based Formation of Cognitive Maps of Semantic Spaces and the Emergence of Abstract Concepts
• On the Transformation of Latent Space in Fine-Tuned NLP Models
• Efficient Methods for Natural Language Processing: A Survey
• Same Neurons, Different Languages: Probing Morphosyntax in Multilingual Pre-trained Models
• Language Modeling, Lexical Translation, Reordering: The Training Process of NMT through the Lens of Classical SMT
• Knowledge Neurons in Pretrained Transformers
• Abstraction and analogy‐making in artificial intelligence
• Transformer Feed-Forward Layers Are Key-Value Memories
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Captum: A unified and generic model interpretability library for PyTorch
• Efficient Transformers: A Survey
• Big Bird: Transformers for Longer Sequences
• Hopfield Networks is All You Need
• Evaluating and Aggregating Feature-based Model Explanations
• Generating Fact Checking Explanations
• Reformer: The Efficient Transformer
• What Does BERT Look At? An Analysis of BERT's Attention
• Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned
• Are Sixteen Heads Really Better than One?
• Generating Long Sequences with Sparse Transformers
• Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context
• Machine learning for neural decoding
• Attention Is All You Need
• Effective Approaches to Attention-based Neural Machine Translation
• Neural Machine Translation by Jointly Learning to Align and Translate
• Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation
• Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation
• Efficient Estimation of Word Representations in Vector Space
• Attention and the detection of signals.
• Orienting of Attention
• Scripts in memory for text
• Features of similarity.
• Some Experiments on the Recognition of Speech, with One and with Two Ears
• Learning Visual-Semantic Subspace Representations
• Explaining Explainability: Recommendations for Effective Use of Concept Activation Vectors
• Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
• Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
• A feature-integration theory of attention

Related

• Intra-neuronal attention within language models Relationships between activation and semantics
• The Process of Categorical Clipping at the Core of the Genesis of Concepts in Synthetic Neural Cognition
• Synthetic Categorical Restructuring large Or How AIs Gradually Extract Efficient Regularities from Their Experience of the World
• Disentangling Neuron Representations with Concept Vectors
• How Do Artificial Intelligences Think? The Three Mathematico-Cognitive Factors of Categorical Segmentation Operated by Synthetic Neurons