AI Paper Atlas

Research map / Architectures and efficiency / Neural architecture search

Recon: Reducing Conflicting Gradients From the Root For Multi-Task Learning

Guangyuan SHI, Qimai Li, Wenlong Zhang, Jiaxin Chen, Xiao-Ming Wu

ICLR 2023 · Architectures and efficiency › Neural architecture search · approach: AutoML pipeline search · 11 citations (OpenAlex)

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
Make layers with the strongest gradient conflicts task-specificMulti-task learning, including PCGrad · Multi-Fashion+MNISTT2 accuracy: 88.68 -> 89.77Improved
Apply Recon to CAGradCAGrad · NYUv2Δm%: 3.74 -> 5.80Improved
Apply Recon to GraddropGraddrop · CelebAAverage Error: 8.33 -> 8.20Improved
Apply Recon to BMTASBMTAS · PASCAL-Context 4-taskΔm%: -2.89 -> -0.19Improved
Choose task-specific layers randomly instead of by conflict score (ablation)Recon · Multi-Fashion+MNISTsevere-conflict reduction: 69.82 -> -52.94Hurt
Choose a similar number of parameters randomly instead of by conflict score (ablation)Recon · Multi-Fashion+MNISTsevere-conflict reduction: 69.82 -> 40.13Hurt
Make the first K layers task-specific rather than conflict-ranked layers (ablation)Recon with CAGrad · Multi-Fashion+MNISTT1 accuracy: 98.28 -> 97.63Hurt
Make the last K layers task-specific rather than conflict-ranked layers (ablation)Recon with CAGrad · Multi-Fashion+MNISTT1 accuracy: 98.28 -> 98.21No clear effect
Select too few layers (ablation)Recon · Multi-Fashion+MNISTCAGrad performanceHurt
Select too many layers (ablation)Recon · Multi-Fashion+MNISTCAGrad performanceNo clear effect
Use a high conflict threshold such as S=0.0 (ablation)Recon · Multi-Fashion+MNISTCAGrad performanceHurt
Use a very low conflict threshold such as S=-0.15 (ablation)Recon · Multi-Fashion+MNISTCAGrad performanceHurt

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