Training vision transformers with spatial loss improves causal circuit grouping
Topographic Training Concentrates Causal Circuits Without Improving Neuron Monosemanticity
Computer Vision and Pattern RecognitionMachine Learning
Summary
Vision transformer models often mix many ideas within single neurons, making them hard to interpret. The authors tried a new training method called TopoLoss that encourages neurons near each other to work on related parts of the task. They found this makes groups of neurons more meaningful and causally effective together, but individual neurons still represent mixed concepts. This shows that looking at individual neurons may miss important patterns in how the model works.
What this means in practice
- •For ai model developers: Improve training methods to create spatially organized neuron groups that better capture causal roles in vision transformer models.
- •For ai interpretability engineers: Use topographic training loss to enhance the effectiveness of existing mechanistic interpretability tools on vision transformers.
Tested on one dataset.
Authors
Gautam Ranka, Shubham Santosh Pandere, Aiden Dsouza
Abstract
Mechanistic interpretability of vision transformers seeks to decompose model computation into human-readable units, but learned representations entangle many concepts in each neuron. Feature superposition is widely treated as the central obstacle to this decomposition, yet most mitigations (sparse autoencoders, dictionary learning) are post-hoc and leave the underlying network unchanged. We ask whether a spatial-locality training loss (TopoLoss) can act as a lightweight, training-time prior that improves interpretability of standard mech-interp tools. Training ViT on ImageNet-100 across multiple TopoLoss weights $α$, we measure causal sufficiency of topographic clusters via activation patching and feature geometry via sparse autoencoders fit to the same residual stream. At $α=1.0$, topographic clusters are 2.79$\times$ more causally sufficient than random unit sets of the same size, with the effect increasing monotonically in $α$. SAE L0 sparsity decreases by 11% and dead-feature fraction rises 19-fold, yet standard neuron-level monosemanticity scores are unchanged, indicating that topographic pressure acts at circuit level, concentrating causal mass into spatially local structures without disentangling individual neurons. This dissociation suggests current neuron-level monosemanticity metrics are insensitive to a class of real interpretability gains, and positions cheap architectural priors as a viable training-time complement to post-hoc tooling.