Importance-Aware Feature Sparsification for Wireless Split Learning
ICS sparsifies wireless split-learning features with server-side Grad-CAM scores, cutting communication cost.
Wireless split learning offloads upper layers, but sending high-dimensional intermediate features remains a communication bottleneck. Importance-aware class-balanced sparsification has the server rank channels with Grad-CAM scores from the true-class logit; clients then keep the top-N channels next round without extra local passes. A non-asymptotic bound relates sparsification ratio and batch size to convergence under a fixed budget. ICS extends to parallel split learning and transformers and outperforms baselines, especially under severe non-i.i.d. data.
- Server ranks channels using Grad-CAM true-class scores
- Class-balanced aggregation limits head-class bias under skew
- Clients reuse the vector with no extra local passes
- Extended to parallel split learning and transformers
- Larger gains under severe non-i.i.d. partitions
From the original article
1 min read at arxiv.orgWireless split learning (SL) reduces on-device computation by offloading upper layers to a server, yet transmitting high-dimensional intermediate features at each iteration remains a major communication bottleneck.
By Bumjun Kim · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.