BMVC 2026

When Neural Collapse Fails to Route

Routing drift in continual learning, and Reliability-Calibrated Routing for progressive ETF classifiers.

Nicolas Larue
ETIS, CY Cergy Paris Universite, ENSEA, CNRS, France
Task-by-task diagnostic curves showing routing drift in ProNC.
Class-IL accuracy follows route accuracy rather than Task-IL accuracy, revealing a cross-task score comparability failure under the ETF classifier.

Takeaways

High Task-IL accuracy does not imply good Class-IL accuracy. Features can still distinguish classes within each task while the largest global score belongs to another task block.

5/6best Class-IL FAA settings
6/6lower Class-IL forgetting settings
+8.03Class FAA in paired eval
+8.62route accuracy in paired eval
Question
Why can an ETF continual classifier keep high Task-IL accuracy while losing Class-IL accuracy?
Finding
The dominant late error is often wrong task-block routing, not within-task class confusion.
Method
RCR keeps the ETF classifier fixed and adds transported, radius-calibrated class reliability evidence.
Outcome
Best reported Class-IL FAA on five of six dataset/buffer settings, with lower forgetting on all six.

Abstract

Progressive neural-collapse classifiers can learn strong task-local decision boundaries while failing the global Class-IL decision. We identify this failure as routing drift: a held-out sample can remain separable within its true task block, yet receive its largest global score from another block.

We formalize route accuracy directly from the global class argmax, separating within-task discrimination from cross-task score comparability. This decomposition motivates Reliability-Calibrated Routing (RCR), a classifier-level correction that keeps the trained backbone and global ETF frame fixed while adding class-local reliability evidence around transported, NC-shrunk class representatives.

Across three datasets and two fixed replay-buffer budgets, RCR gives the best Class-IL FAA on five of six settings and reduces Class-IL forgetting on all six, with the strongest gains in harder routing regimes.

Routing Drift

Problem. Class-IL accuracy is a product of two decisions, but standard reporting usually exposes only the final product. This can make a model look like it forgot the old classes, even when most of the remaining error is a cross-task routing error.

Class-IL correctness requires two events: the predicted class must fall in the correct task block, and the within-block class decision must be correct. In progressive ETF classifiers, the second event can remain strong while the first collapses.

$\mathrm{ClassIL}(x)=\mathbf{1}\{\hat{\tau}(x)=\tau(y)\}\,\mathbf{1}\{\hat{y}(x)=y\mid\hat{\tau}(x)=\tau(y)\}.$

Route accuracy measures whether the global argmax lands in the correct task block. It is read from the classifier output; it is not a learned task oracle.

Global prediction $\hat y(x)=\arg\max_{c\in\mathcal C_t} s(x,c),\qquad \hat\tau(x)=\tau(\hat y(x)).$
Route accuracy $\mathrm{RouteAcc}=\mathbb E_{(x,y)}\left[\mathbf{1}\{\hat\tau(x)=\tau(y)\}\right].$
Replay blind spot for routing drift.
Small replay buffers can mask routing drift because stored samples are repeatedly optimized and often remain better routed than held-out samples.

Reliability-Calibrated Routing

RCR corrects the classifier while keeping the trained backbone and ETF weights fixed. It does not require retraining or a learned task router.

Each class receives two scores: alignment with its ETF vertex, and radial plausibility around its transported prototype. The second score accounts for differences in class dispersion after feature drift.

ETF frame

Use the original ETF score as global neural-collapse evidence.

Transport

Move class representatives into the current feature frame before scoring old classes.

Reliability

Add radial plausibility with competitive radii and NC-consistent shrinkage.

ETF evidence $s_{\mathrm{ETF}}(x,c)=\langle z(x),u_c\rangle,$ where $u_c$ is the current ETF vertex.
Transported, NC-shrunk representative $p_c^t=\mathrm{norm}\!\left(A_t p_c^{t-1}\right),\qquad \tilde p_c=\mathrm{norm}\!\left(\alpha_c p_c^t+(1-\alpha_c)u_c\right).$
Radial plausibility $d_{\cos}(a,b)=2-2\langle a,b\rangle,\qquad s_{\mathrm{rad}}(x,c)=-\,d_{\cos}(z(x),\tilde p_c)/(\rho_c+\varepsilon).$
RCR score $s_{\mathrm{RCR}}(x,c)=(1-\omega)\,\mathrm{Std}_c(s_{\mathrm{ETF}}(x,c))+ \omega\,\mathrm{Std}_c(s_{\mathrm{rad}}(x,c)),\qquad \omega=\frac{q}{q+1},$ where $q$ is the number of classes per task.
Keep the trained model fixed.
RCR is evaluated on the same end-of-task backbone and ETF classifier.
Transport class representatives.
Old class centers are moved into the current feature frame before nearest-prototype evidence is computed.
Estimate local reliability.
Competitive radii encode how much radial evidence each class should contribute.
Compose comparable scores.
The ETF and radial scores are standardized per sample and mixed with a fixed rule.

When class centers coincide with ETF vertices and all radii agree, RCR makes the same class predictions as the ETF classifier. Its correction addresses deviations from this ideal geometry.

ETF-only view of a corrected routing case. RCR-corrected view of the same routing case.
The same feature-space case under ETF-only evidence and RCR evidence. The correction adds local reliability around transported class representatives while preserving the ETF frame.

Results

The paired comparison below evaluates $s_{\mathrm{ETF}}$ and $s_{\mathrm{RCR}}$ on the same frozen checkpoints over seeds 0, 1, and 2. RCR improves Class-IL and route accuracy together, while Task-IL remains stable.

The largest gains occur on CIFAR-100 and TinyImageNet, where the Task-IL / route gap is widest. CIFAR-10 with buffer 500 is already close to the neural-collapse ideal, so the gain is smaller; this is consistent with RCR being inactive when ETF geometry is already reliable.

Dataset / buffer 200 Score Class FAA ↑ Class FF ↓ Task FAA ↑ Route ↑
CIFAR-10ETF66.1431.4796.7466.58
CIFAR-10RCR70.36 (+4.22)22.85 (-8.62)96.6770.82 (+4.24)
CIFAR-100ETF41.7138.5585.5342.81
CIFAR-100RCR49.74 (+8.03)16.27 (-22.28)87.2851.43 (+8.62)
TinyImageNetETF25.7743.4568.3028.58
TinyImageNetRCR31.42 (+5.65)23.84 (-19.61)69.5934.76 (+6.18)
Dataset / buffer 500 Score Class FAA ↑ Class FF ↓ Task FAA ↑ Route ↑
CIFAR-10ETF72.1624.1896.8272.71
CIFAR-10RCR73.68 (+1.52)17.73 (-6.45)96.8174.33 (+1.62)
CIFAR-100ETF48.6523.1685.5450.14
CIFAR-100RCR53.27 (+4.62)10.18 (-12.98)87.7455.04 (+4.90)
TinyImageNetETF26.6141.1667.7129.37
TinyImageNetRCR32.24 (+5.63)22.47 (-18.69)70.1635.46 (+6.09)

FAA: final average accuracy. FF: final forgetting. See the arXiv paper for full comparisons and ablations.

Reproducibility

The GitHub release archives include the per-task features needed to recompute accuracy, forgetting, and route accuracy without retraining. Each archive covers three seeds for one dataset and replay budget. Follow the evaluation commands in the README.

Citation

@inproceedings{larue2026rcr,
  title     = {When Neural Collapse Fails to Route: Routing Drift in Continual Learning},
  author    = {Larue, Nicolas},
  booktitle = {British Machine Vision Conference (BMVC)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2611.22429}
}