TALOR
A plug-and-play feature correction method for open-set test-time adaptation (OSTTA). Based on the geometric structure of the source model's linear classification head, it decomposes principal and tail coordinates and estimates drift via a closed-form weighted ridge regression, requiring no backpropagation or memory bank.
Timeline (last 90 days)
TALOR is an OSTTA correction module requiring neither backpropagation nor a memory bank, using the geometry of the source model's linear classification head for feature correction.
TALOR uses the basis induced by the source linear head weights to decompose normalized features into high-gain principal coordinates and low-gain tail coordinates.
TALOR estimates, in the principal-coordinate subspace, a soft-routed affine correction consisting of a regime-level principal bias and a slope term conditioned on the tail coordinates.
TALOR uses weighted ridge regression to solve in closed form for the regime-level drift from the principal coordinates, with weights derivable from tail-coordinate statistics.
In evaluations on the ImageNet-C benchmark with six csOOD datasets, TALOR's H-score is 2.2% higher than the next-best method UniEnt.
In the same evaluation, TALOR runs 2.5 times faster than UniEnt and uses only 12% of UniEnt's GPU memory.
As a post-correction plugin, TALOR raises the H-score of the COME method to 68.4%, with only 2% additional memory overhead.
TALOR is not tied to a specific backbone network or adaptation strategy and can be layered onto existing OSTTA pipelines as a general component.
TALOR's effectiveness may depend on the quality of the source model's linear classification head; whether the information capacity of the tail coordinates is sufficient in more complex open-set scenarios still requires verification.
All Facts (9)
TALOR is an OSTTA correction module requiring neither backpropagation nor a memory bank, using the geometry of the source model's linear classification head for feature correction.
50%UnverifiedTALOR uses the basis induced by the source linear head weights to decompose normalized features into high-gain principal coordinates and low-gain tail coordinates.
50%UnverifiedTALOR estimates, in the principal-coordinate subspace, a soft-routed affine correction consisting of a regime-level principal bias and a slope term conditioned on the tail coordinates.
50%UnverifiedTALOR uses weighted ridge regression to solve in closed form for the regime-level drift from the principal coordinates, with weights derivable from tail-coordinate statistics.
50%UnverifiedIn evaluations on the ImageNet-C benchmark with six csOOD datasets, TALOR's H-score is 2.2% higher than the next-best method UniEnt.
50%UnverifiedIn the same evaluation, TALOR runs 2.5 times faster than UniEnt and uses only 12% of UniEnt's GPU memory.
50%UnverifiedAs a post-correction plugin, TALOR raises the H-score of the COME method to 68.4%, with only 2% additional memory overhead.
50%UnverifiedTALOR is not tied to a specific backbone network or adaptation strategy and can be layered onto existing OSTTA pipelines as a general component.
50%UnverifiedTALOR's effectiveness may depend on the quality of the source model's linear classification head; whether the information capacity of the tail coordinates is sufficient in more complex open-set scenarios still requires verification.
50%