Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Ferrao, Jeremias, Mikaelson, Luhan, Pepper, Keenan, Antolin, Natalia Perez-Campanero
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2505.00509
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866916716037013504
author Ferrao, Jeremias
Mikaelson, Luhan
Pepper, Keenan
Antolin, Natalia Perez-Campanero
author_facet Ferrao, Jeremias
Mikaelson, Luhan
Pepper, Keenan
Antolin, Natalia Perez-Campanero
contents A growing intuition in machine learning suggests a link between sparsity and interpretability. We introduce a novel self-ablation mechanism to investigate this connection ante-hoc in the context of language transformers. Our approach dynamically enforces a k-winner-takes-all constraint, forcing the model to demonstrate selective activation across neuron and attention units. Unlike post-hoc methods that analyze already-trained models, our approach integrates interpretability directly into model training, promoting feature localization from inception. Training small models on the TinyStories dataset and employing interpretability tests, we find that self-ablation leads to more localized circuits, concentrated feature representations, and increased neuron specialization without compromising language modelling performance. Surprisingly, our method also decreased overall sparsity, indicating that self-ablation promotes specialization rather than widespread inactivity. This reveals a complex interplay between sparsity and interpretability, where decreased global sparsity can coexist with increased local specialization, leading to enhanced interpretability. To facilitate reproducibility, we make our code available at https://github.com/keenanpepper/self-ablating-transformers.
format Preprint
id arxiv_https___arxiv_org_abs_2505_00509
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Self-Ablating Transformers: More Interpretability, Less Sparsity
Ferrao, Jeremias
Mikaelson, Luhan
Pepper, Keenan
Antolin, Natalia Perez-Campanero
Machine Learning
A growing intuition in machine learning suggests a link between sparsity and interpretability. We introduce a novel self-ablation mechanism to investigate this connection ante-hoc in the context of language transformers. Our approach dynamically enforces a k-winner-takes-all constraint, forcing the model to demonstrate selective activation across neuron and attention units. Unlike post-hoc methods that analyze already-trained models, our approach integrates interpretability directly into model training, promoting feature localization from inception. Training small models on the TinyStories dataset and employing interpretability tests, we find that self-ablation leads to more localized circuits, concentrated feature representations, and increased neuron specialization without compromising language modelling performance. Surprisingly, our method also decreased overall sparsity, indicating that self-ablation promotes specialization rather than widespread inactivity. This reveals a complex interplay between sparsity and interpretability, where decreased global sparsity can coexist with increased local specialization, leading to enhanced interpretability. To facilitate reproducibility, we make our code available at https://github.com/keenanpepper/self-ablating-transformers.
title Self-Ablating Transformers: More Interpretability, Less Sparsity
topic Machine Learning
url https://arxiv.org/abs/2505.00509