When Neural Collapse Fails to Route: Routing Drift in Continual Learning
Organizations: ETIS, CY Cergy Paris Université ENSEA, CNRS, France
Abstract
Progressive neural-collapse classifiers can learn strong task-local decision boundaries while failing the global Class-IL decision. We identify the failure as routing drift: under an expanding equiangular tight frame (ETF) classifier, a held-out sample can remain separable within its true task block, yet receive its largest global score from another block. This produces high Task-IL accuracy but low Class-IL accuracy, and small replay buffers can mask the problem because stored samples are optimized repeatedly and often remain well routed. We formalize route accuracy directly from the global class argmax and use it to separate within-task discrimination from cross-task score comparability. This decomposition suggests a classifier-level correction: preserve the global ETF frame, but add class-local reliability evidence when transported class measures deviate from the neural-collapse ideal. We develop Reliability-Calibrated Routing (RCR), a lightweight post-hoc classifier that composes standardized ETF alignment with radial plausibility around transported, NC-shrunk class representatives. Across three datasets and two fixed replay-buffer budgets, this NC-consistent correction gives the best Class-IL FAA on five of six settings and reduces Class-IL forgetting on all six. The largest gains occur in harder routing regimes: up to FAA over the reported reference, and up to FAA and final route accuracy in paired evaluations, while preserving Task-IL performance. These results position routing drift as a measurable failure mode of ETF-based continual learning and show that reliability-calibrated class evidence can repair part of the lost global routing without retraining the representation. Code and evaluation artifacts are available at https://towzeur.github.io/RCR/.
Figures & tables
| Dataset | Buffer | Score | Class-IL | Task-IL | Route | ||
|---|---|---|---|---|---|---|---|
| FAA | FF | FAA | FF | ||||
| CIFAR-10 | 200 | 66.14 | 31.47 | 96.74 | 0.62 | 66.58 | |
| (Ours) | 70.36 (+4.22) | 22.85 (-8.62) | 96.67 (-0.07) | 0.73 (+0.11) | 70.82 (+4.24) | ||
| 500 | 72.16 | 24.18 | 96.82 | 0.47 | 72.71 | ||
| (Ours) | 73.68 (+1.52) | 17.73 (-6.45) | 96.81 (-0.01) | 0.54 (+0.07) | 74.33 (+1.62) | ||
| CIFAR-100 | 200 | 41.71 | 38.55 | 85.53 | 4.51 | 42.81 | |
| Setting | Method | Venue | Seq-CIFAR-10 | Seq-CIFAR-100 | Seq-TinyImageNet | |||
|---|---|---|---|---|---|---|---|---|
| FAA | FF | FAA | FF | FAA | FF | |||
| Class-IL | ER ( Riemer et al.(2019)Riemer, Cases, Ajemian, Liu, Rish, Tu, and Tesauro ) | ICLR 2019 | ||||||
| iCaRL \citep rebuffi2017icarl | CVPR 2017 | 31.06 1.91 | ||||||
| DER \citep buzzega2020dark | NeurIPS 2020 | |||||||
| DER++ \citep buzzega2020dark | NeurIPS 2020 | |||||||
| Co 2 L \citep cha2021co2l | ICCV 2021 | |||||||
| Setting | Method | Venue | Seq-CIFAR-10 | Seq-CIFAR-100 | Seq-TinyImageNet | |||
|---|---|---|---|---|---|---|---|---|
| FAA | FF | FAA | FF | FAA | FF | |||
| Class-IL | ER \citep riemer2019learning | ICLR 2019 | ||||||
| iCaRL \citep rebuffi2017icarl | CVPR 2017 | 37.30 1.42 | ||||||
| DER \citep buzzega2020dark | NeurIPS 2020 | |||||||
| DER++ \citep buzzega2020dark | NeurIPS 2020 | |||||||
| Co 2 L \citep cha2021co2l | ICCV 2021 | |||||||
| Head | ETF | Transport | Radial | Comp. | NC shrink | Class-IL | Route |
|---|---|---|---|---|---|---|---|
| ETF only | ✓ | ||||||
| Radial only | ✓ | ✓ | ✓ | ||||
| No transport | ✓ | ✓ | ✓ | ||||
| Equal radius | ✓ | ✓ | ✓ | ||||
| No NC shrink | ✓ | ✓ | ✓ | ✓ | |||
| Ours (RCR) | ✓ | ✓ | ✓ | ✓ | ✓ |
| Head | Class-IL | Task-IL | Route | Wrong-task |
|---|---|---|---|---|
| ProNC ETF | ||||
| Replay task-bias | ||||
| Ours (RCR) |
| Dataset | Buf. | NC1 disp. | CV | Task–Route gap | Route | Class | |
|---|---|---|---|---|---|---|---|
| Seq-CIFAR-10 | 200 | ||||||
| Seq-CIFAR-10 | 500 | ||||||
| Seq-CIFAR-100 | 200 | ||||||
| Seq-CIFAR-100 | 500 | ||||||
| Seq-TinyImageNet | 200 | ||||||
| Seq-TinyImageNet | 500 |
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Buffer | Epochs | LR | Milestones | ||
|---|---|---|---|---|---|---|
| Seq-CIFAR-10 | 200 | 50 | 0.01 | 35, 45 | 13 | 90 |
| Seq-CIFAR-10 | 500 | 50 | 0.01 | 35, 45 | 12 | 80 |
| Seq-CIFAR-100 | 200 | 50 | 0.03 | 35, 45 | 18 | 170 |
| Seq-CIFAR-100 | 500 | 50 | 0.03 | 35, 45 | 18 | 170 |
| Seq-TinyImageNet | 200 | 100 | 0.03 | 45, 75, 95 | 20 | 165 |
| Seq-TinyImageNet | 500 | 100 | 0.03 | 45, 75, 95 | 20 | 165 |
| Type | Symbol | Description |
| Protocol and indexing | ||
| Set | Ordered continual-learning task sequence. | |
| Index | Current training or evaluation stage. | |
| Set | Training data introduced by task . | |
| Set | Classes observed after task . | |
| Scalar | Number of classes per task block. | |
| Dataset | Buffer | Head | Class FAA | Class FF | Task FAA | Task FF | Route | |
|---|---|---|---|---|---|---|---|---|
| CIFAR-10 | 200 | ETF reference | 3 | 66.14 1.69 | 31.47 2.46 | 96.74 0.12 | 0.62 0.20 | 66.58 1.64 |
| CE head (audit) | 3 | 54.63 0.15 | 52.99 0.03 | 96.62 0.24 | 0.75 0.24 | 55.04 0.09 | ||
| Transported NCM | 3 | 63.46 0.75 | 20.75 1.21 | 96.11 0.53 | 1.36 0.66 | 63.98 0.66 | ||
| Radial-only | 3 | 64.38 1.23 | 15.80 1.84 | 96.11 0.53 | 1.36 0.66 | 65.06 0.98 | ||
| CompMix | 3 | 69.84 0.76 | 24.39 1.40 | 96.63 0.10 | 0.75 0.22 | 70.30 0.80 | ||
| CompMix-Fixed | 3 | 69.84 0.76 | 24.15 1.46 | 96.63 0.10 | 0.77 0.22 | 70.30 0.80 |
| Head | Class FAA | Class FF | Task FAA | Task FF | Route |
|---|---|---|---|---|---|
| NCT-style ETF | 31.62 | 27.05 | 79.93 | 4.99 | 33.58 |
| NCT-style ETF + RCR-style ramp | 34.02 | 25.84 | 81.33 | 3.31 | 35.89 |
| Dataset/Buffer | Head | Class-IL | Task-IL | Route |
|---|---|---|---|---|
| Seq-CIFAR-10 / 200 | ETF only | |||
| Transported NCM | ||||
| Radial only | ||||
| ETF+NCM | ||||
| ETF+raw radial | ||||
| ETF+equal radius |
| Set | Class | Task | Route | |
|---|---|---|---|---|
| C10/200 | 0.00 | |||
| 0.25 | ||||
| 0.50 | ||||
| 0.75 | ||||
| Ours | ||||
| 1.00 |
| Buffer | Mixture rule | Class FAA | Class FF | Task FAA | Route | |
|---|---|---|---|---|---|---|
| 200 | Stage schedule | 2 | 49.55 | 19.31 | 87.49 | 51.21 |
| 200 | Ours: fixed cap | 2 | 49.49 | 16.43 | 87.39 | 51.18 |
| 500 | Stage schedule | 2 | 53.12 | 11.44 | 87.72 | 54.91 |
| 500 | Ours: fixed cap | 2 | 53.13 | 10.29 | 87.71 | 54.94 |
| Class-IL | Task-IL | Route | Wrong-task | |
|---|---|---|---|---|
| 0.00 | ||||
| 0.10 | ||||
| Ours: 0.25 | ||||
| 0.50 | ||||
| 1.00 |
| Dataset / Buffer | Setting | Class FAA | Class FF | |
|---|---|---|---|---|
| Seq-CIFAR-10 / 200 | No upper safeguard | 1.00 | ||
| Seq-CIFAR-10 / 200 | Ours | 0.95 | ||
| Seq-CIFAR-10 / 500 | No upper safeguard | 1.00 | ||
| Seq-CIFAR-10 / 500 | Ours | 0.95 | ||
| Seq-CIFAR-100 / 200 | No upper safeguard | 1.00 | ||
| Seq-CIFAR-100 / 200 | Ours | 0.95 |
| Dataset/Buffer | Head | Class-IL | Task-IL | Route | Wrong-task |
|---|---|---|---|---|---|
| Seq-CIFAR-10 / 200 | ProNC ETF | ||||
| Replay task-bias | |||||
| Replay class-bias | |||||
| Replay task affine | |||||
| Ours (RCR) | |||||
| Seq-CIFAR-10 / 500 | ProNC ETF |
| Stage | NC1 disp. | CV | Route | Class-IL | |
|---|---|---|---|---|---|
| T3 | |||||
| T10 |
| Dataset/Buffer | Head | Class-IL | Task-IL | Route |
|---|---|---|---|---|
| Seq-CIFAR-10 / 200 | ProNC ETF | |||
| RCR | ||||
| ProNC + true route | ||||
| RCR + true route | ||||
| Seq-CIFAR-10 / 500 | ProNC ETF | |||
| RCR |
| Dataset/Buffer | Signal | Route error | Class |
|---|---|---|---|
| Seq-CIFAR-10 / 200 | |||
| NC1 disp. | |||
| Age | |||
| Seq-CIFAR-10 / 500 | |||
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 174 papers