AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Enabling Fine-Tuning of Direct Feedback Alignment via Feedback-Weight Matching

Yunseok Lee, Seulki Lee

ICLR 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply feedback-weight matching before DFA fine-tuning6-layer FC network · SVHN (from CIFAR-100)test accuracy: 74.70 -> 82.67Improved
apply feedback-weight matching before DFA fine-tuningBERT-Small · GLUE STSBPearson correlation: 0.10 -> 0.76Improved
apply feedback-weight matching before DFA fine-tuningBERT-Small · GLUE CoLAMatthews correlation: 0.06 -> 0.53Improved
apply feedback-weight matching before DFA fine-tuningViT-Small · ImageNetteaccuracy: 0.210 -> 0.319Improved
apply feedback-weight matching before DFA fine-tuning4-layer FC network · CIFAR-10 (from TinyImageNet)test accuracy: 56.75 -> 55.51Mixed
apply feedback-weight matching before DFA fine-tuningBERT-Small · GLUE QQPaccuracy: 67.0 -> 65.8Hurt
add weight matching (re-initialize weights from feedback) (ablation)DFA fine-tuning · SVHN (4-layer, from TinyImageNet)test accuracy: 79.77 -> 83.16Improved
add feedback matching (decompose pre-trained weights) (ablation)DFA fine-tuning · CIFAR-10 (6-layer, from TinyImageNet)test accuracy: 55.03 -> 55.54Improved
add weight decay on top of feedback-weight matching (ablation)4-layer FC network · CIFAR-10 (from CIFAR-100)test accuracy: 48.82 -> 55.38Improved
add weight decay to standard DFA without matching (ablation)standard DFA fine-tuning · CIFAR-10 (4-layer, from TinyImageNet)test accuracyNo clear effect
train feedback matrices 3 epochs instead of 1 (ablation)feedback matching · CIFAR-10 (4-layer, from CIFAR-100)test accuracy: 51.38 -> 55.38Improved
apply strong weight decay (1e-2) to feedback-weight matching (ablation)4-layer FC network · CIFAR-10 (from CIFAR-100)train/test accuracyHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy