Diagnosing Prototype Estimation in Few-Shot Class-Incremental Learning: Training-Free Geometry Selection and Unlabeled-Query Prototype Rectification
Organizations: Islamic University of Technology
Abstract
Frozen visual encoders make prototype heads competitive in few-shot class-incremental learning (FSCIL), but it remains unclear how much of the residual head gap is sensitive to noisy novel-class means rather than to the subsequent changes introduced by classifier refitting. On identical features, we construct a retrospective matched counterfactual: first replace each five-shot novel mean with its complete-data counterpart under fixed cosine scoring, then fit a linear probe. Across sixteen strong encoder dataset pairs, mean replacement accounts for the majority of this ordered gap. Shot and classwise diagnostics support a finite-sample mechanism, whereas a weaker base-only encoder does not satisfy the same criterion. This diagnosis motivates ShrinkSel, a support-only selector over prototype radius and covariance shrinkage whose family contains shared LDA as an exact fallback. Across sixteen evaluated pairs, it improves LDA on fifteen by 0.04-1.52 points and selects the exact LDA endpoint on Aircraft. We further introduce Query Prototype Rectification (QPR), a batch-transductive extension that uses gated unlabeled queries to update novel directions while preserving the selected geometry. On CLIP ViT-B/16, QPR adds 1.21-2.03 points; exact query-wise self-exclusion retains the benefit, showing that the gains do not require self-inclusion. The update repairs scarce novel prototypes at a measurable base-class cost and weakens with plentiful supports or very small cohorts. Thus, the evidence available for prototype estimation is a central constraint in frozen-feature FSCIL.
Figures & tables
| Method | Accuracy in each session (%) | Avg. | PD | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | |||
| Cosine-NCM | 81.63 | 79.14 | 77.49 | 74.25 | 73.26 | 71.69 | 70.81 | 70.04 | 68.53 | 68.72 | 68.48 | 73.09 | 13.16 |
| TEEN (released) | 81.63 | 78.54 | 76.64 | 72.90 | 71.85 | 70.33 | 69.50 | 68.68 | 67.08 | 67.26 | 67.09 | 71.95 | 14.54 |
| ULA (released) | 81.63 | 79.18 | 77.57 | 74.30 | 73.20 | 71.63 | 70.93 | 70.16 | 68.61 | 68.83 | 68.62 | 73.15 | 13.01 |
| FeCAM | 77.06 | 71.92 | 67.35 | 63.21 | 60.64 | 58.29 | 55.68 | 53.69 | 51.68 | 50.81 | 49.71 | 60.00 | 27.35 |
| HyCal image-only reduction † | 79.78 | 74.06 | 69.27 | 64.92 | 62.10 | 59.66 | 56.85 | 54.87 | 52.70 | 51.95 | 50.80 | 61.54 | 28.98 |
| Method | CUB-200 | CIFAR-100 | miniImageNet | |||
|---|---|---|---|---|---|---|
| Avg. | Parent [95%] | Avg. | Parent [95%] | Avg. | Parent [95%] | |
| Reference head: Cosine-NCM | ||||||
| Support-only | 73.09 | — | 67.88 | — | 89.00 | — |
| QPR | 73.76 | +0.67 [+0.60, +0.73] | 68.25 | +0.37 [+0.32, +0.42] | 89.54 | +0.54 [+0.51, +0.58] |
| QPR (self-excl.) | 73.83 | +0.73 [+0.68, +0.79] | 68.24 | +0.36 [+0.31, +0.41] | 89.54 | +0.54 [+0.51, +0.58] |
| TIM-GD | 72.99 | -0.11 [-0.22, 0.00] | 69.40 | +1.51 [+1.43, +1.60] | 90.33 | +1.33 [+1.28, +1.38] |
Class-Incremental Learning
Class-incremental learning studies how models can add new categories over time while retaining the ability to recognize previously learned classes.
See all 171 papers