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

We revisit and extend model stitching (Lenc & Vedaldi 2015) as a methodologyto study the internal representations of neural networks. Given two trained andfrozen models $A$ and $B$, we consider a "stitched model'' formed by connectingthe bottom-layers of $A$ to the top-layers of $B$, with a simple trainablelayer between them. We argue that model stitching is a powerful and perhapsunder-appreciated tool, which reveals aspects of representations that measuressuch as centered kernel alignment (CKA) cannot. Through extensive experiments,we use model stitching to obtain quantitative verifications for intuitivestatements such as "good networks learn similar representations'', bydemonstrating that good networks of the same architecture, but trained in verydifferent ways (e.g.: supervised vs. self-supervised learning), can be stitchedto each other without drop in performance. We also give evidence for theintuition that "more is better'' by showing that representations learnt with(1) more data, (2) bigger width, or (3) more training time can be "plugged in''to weaker models to improve performance. Finally, our experiments reveal a newstructural property of SGD which we call "stitching connectivity'', akin tomode-connectivity: typical minima reached by SGD can all be stitched to eachother with minimal change in accuracy.

Revisiting Model Stitching to Compare Neural Representations