English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

In light of the recent widespread adoption of AI systems, understanding theinternal information processing of neural networks has become increasinglycritical. Most recently, machine vision has seen remarkable progress by scalingneural networks to unprecedented levels in dataset and model size. We here askwhether this extraordinary increase in scale also positively impacts the fieldof mechanistic interpretability. In other words, has our understanding of theinner workings of scaled neural networks improved as well? We here use apsychophysical paradigm to quantify mechanistic interpretability for a diversesuite of models and find no scaling effect for interpretability - neither formodel nor dataset size. Specifically, none of the nine investigatedstate-of-the-art models are easier to interpret than the GoogLeNet model fromalmost a decade ago. Latest-generation vision models appear even lessinterpretable than older architectures, hinting at a regression rather thanimprovement, with modern models sacrificing interpretability for accuracy.These results highlight the need for models explicitly designed to bemechanistically interpretable and the need for more helpful interpretabilitymethods to increase our understanding of networks at an atomic level. Werelease a dataset containing more than 120'000 human responses from ourpsychophysical evaluation of 767 units across nine models. This dataset ismeant to facilitate research on automated instead of human-basedinterpretability evaluations that can ultimately be leveraged to directlyoptimize the mechanistic interpretability of models.

Scale Alone Does not Improve Mechanistic Interpretability in Vision Models