Forgetting Is Not Erasure, and One Attention Layer Can Power Generation: Two December 2025 Papers Rethink AI Efficiency

Forum topic · ✨步子哥 · 2026-01-10

Summary

A Chinese tech forum post reviews two arXiv papers from December 2025 that challenge the scaling-law orthodoxy in AI. First, ETH Zurich researchers mathematically decompose catastrophic forgetting into shallow forgetting (output-layer accuracy loss caused by classifier misalignment and statistical artifacts) and deep forgetting (geometric drift of old-task features into subspaces orthogonal to active directions). Using Neural Collapse geometry, they show features remain linearly separable beneath apparent forgetting, and that an arbitrarily small replay buffer suffices to prevent deep forgetting, while shallow forgetting requires much larger buffers. Experiments on CIFAR-100 and MiniImageNet validate this asymmetry. Second, Apple's Feature Auto-Encoder (FAE) demonstrates that a single self-attention layer with linear projections, attached to frozen pretrained visual encoders like DINOv2 and SigLIP, can convert understanding-oriented encoders into generative engines. FAE achieves 1.29 FID on ImageNet 256x256 with classifier-free guidance, 7-13x faster convergence than baselines, and strong results on MS-COCO and flow-matching models. The post frames both works as evidence for a modular future: frozen universal backbones plus lightweight adapters, reducing compute, cost, and barriers to entry.

Overview

A forum post on zhichai.net discusses two December 2025 arXiv papers that, together, argue "less is more" in AI: catastrophic forgetting is merely knowledge hiding rather than erasure, and a single attention layer can turn a frozen visual encoder into a generative engine.

Paper 1: Shallow and Deep Forgetting (ETH Zurich)

arXiv:2512.07400 — *Asymptotic Analysis of Shallow and Deep Forgetting in Replay with Neural Collapse*

The paper decomposes catastrophic forgetting into two layers:

Shallow forgetting (output level): classifier-head accuracy on old tasks collapses, measured as $F^{\text{shallow}}_{i \to j} = A_{jj} - A_{ij}$ . Causes include classifier misalignment and statistical artifacts (missing covariance, mean-norm inflation). Repairing it requires substantial replay data, with estimation error converging at $O(b^{-1/2})$ .
Deep forgetting (feature level): measured via frozen linear probes as $F^{\text{deep}}_{i \to j} = A^*_{jj} - A^*_{ij}$ . Surprisingly, features often remain linearly separable even when outputs fail — old knowledge is stored in subspaces orthogonal to the active ones.

Using Neural Collapse geometry, the authors show that without replay, old-task samples' projections on orthogonal directions decay exponentially:

\phi_{t,S^\perp}(x) = (1 - \eta\lambda)^{t - t_0} \phi_{t_0,S^\perp}(x)

turning old samples into OOD samples and unifying continual learning with OOD detection.

Replay efficiency gap

Deep forgetting: an arbitrarily small non-zero replay ratio $$r > 0$$ anchors feature geometry, keeping linear separability SNR bounded long-term: $\lim_{t\to\infty} \text{SNR}(c_1, c_2) \in \Theta(r^2)$ .
Shallow forgetting: requires much larger buffers due to small-buffer statistical bias.

Experiments on CIFAR-100 and MiniImageNet confirm: tiny buffers nearly eliminate deep forgetting, while shallow forgetting demands more data.

Single-head vs multi-head

In class-incremental (single-head) settings, replay induces global Neural Collapse with feature rank up to $$nK - 1$$ ; in task-incremental (multi-head) settings, per-task NC subspaces are mutually orthogonal with rank up to $$n(K-1)$$ — suggesting modular designs naturally reduce interference.

Paper 2: Apple's Feature Auto-Encoder (FAE)

arXiv:2512.07829 — *Adapting Pretrained Visual Encoders for Image Generation*

Instead of training a complex VAE from scratch, FAE attaches a minimal interface to frozen pretrained encoders (DINOv2, SigLIP):

Encoder adapter: a single self-attention layer plus linear projection that de-redundifies patch embeddings
Feature decoder: Transformer reconstructing original embeddings via L2 loss + KL regularization
Pixel decoder: adversarial, perceptual, and reconstruction losses for image generation

The adaptation strength stays deliberately below the pretrained signal, avoiding overfitting and preserving the original semantic geometry.

Results

| Benchmark | FAE FID (no CFG) | FAE FID (with CFG) | Notes | |---|---|---|---| | ImageNet 256×256 (800 epochs) | 1.48 | 1.29 | Beats SD-VAE baselines | | ImageNet 256×256 (80 epochs) | 2.08 | 1.70 | 7–13× faster convergence | | MS-COCO (CC12M data) | 7.47 | 6.90 | Reaches targets with less data | | STARFlow (flow matching) | 2.67 | – | Better than SD-VAE |

A low-dimensional 32-dim latent space stabilizes diffusion, aided by Gaussian noise pretraining and time-shifting tricks.

The bigger picture: a bend in the scaling curve

The post situates both papers against growing doubts about scaling laws: Ilya Sutskever's remarks that scaling is nearing its ceiling, Epoch AI's analysis of barriers through 2030 (power, chips, data), and Harvard's "Scaling Laws for Precision" work on diminishing returns. Both papers offer an alternative path — frozen universal backbones plus lightweight adapters — implying:

Greener computation (no trillion-parameter retraining from scratch)
Lower barriers for small labs
The interface/adapter layer as the new competitive frontier

Scaling laws don't end; the field enters a hybrid era where core backbones grow moderately while peripheral tasks are handled by efficient modular components.

References

1. Asymptotic Analysis of Shallow and Deep Forgetting in Replay with Neural Collapse. arXiv:2512.07400, 2025. 2. Adapting Pretrained Visual Encoders for Image Generation. arXiv:2512.07829, 2025. 3. AI Giants Rethink Model Training Strategy as Scaling Laws Break Down. The Batch, DeepLearning.AI. 4. Can AI scaling continue through 2030? Epoch AI Blog. 5. The End of Scaling Laws: How Harvard's "Scaling Laws for Precision" Revolutionizes LLM Training. Towards AI, 2025.