TALOR rectifies open-set test-time features without backprop
Original title: Tail Cues, Principal Corrections: Plug-and-Play Rectification for Open-Set Test-Time Adaptation
On ImageNet-C with six csOOD datasets it beats UniEnt by 2.2% H-score while using 12% of the GPU memory.
TALOR is a lightweight, backward-free module for open-set test-time adaptation that uses the source linear head as a geometric anchor. It splits normalized features into high-gain principal coordinates and low-gain tail coordinates, estimates an affine drift correction in the principal subspace, and feeds the reconstructed feature to the source head. The authors report a 2.2% H-score gain over UniEnt, 2.5x faster runtime and 12% of its GPU memory. As a plug-in it raises COME to 68.4% H-score with 2% extra memory.
Unconfirmed: Benchmark gains are reported by the paper’s authors only.
- No parameter updates or mutable memories are used at test time.
- Correction uses weighted ridge regression in the principal subspace.
- COME reaches 68.4% H-score with 2% added memory overhead.
- Evaluated on ImageNet-C plus six csOOD datasets.
From the original article
1 min read at arxiv.orgOpen-Set Test-Time Adaptation (OSTTA) operates on non-stationary target streams where covariate and semantic shifts coexist. Existing TTA methods often obtain adaptation signals by updating model parameters or maintaining mutable memories, which can be costly and vulnerable to unknown samples. We present TALOR, a lightweight, backward-free rectification module that uses the source linear head as a geometric anchor.
By Yingkai Yang · 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.