cs.AINov 12, 2026

Model Annealing: Data-Efficient Learning through Review and Consolidation

Authors: Ali Mohammadi Ruzbahani, Rezvan Abouzadeh

Organizations: Smart Cyber-Physical Systems (SCPS) Lab, University of Calgary, Calgary, AB, Canada · MetaKet Inc., Calgary, AB, Canada

Abstract

A small labelled dataset is memorized long before the training budget runs out, and the one-hot labels that remain to be fitted then mostly sharpen confidence. Model annealing spends the rest of the budget on review and consolidation, relearning from the network's own knowledge. Training is divided into episodes whose learning rate is raised again and cooled, every cooled state joins a memory of past selves, and each new episode learns from the labels, from the tempered predictions of this memory, and from adversarial dreams sought where student and memory disagree most. The memory averages predictions rather than weights, and the deployed network is the last student, so inference costs nothing extra. Across CIFAR-100, Tiny-ImageNet and Imagewoof, four architecture families and 5% to 100% of the labels, model annealing raises the top-1 accuracy of the AdamW-trained ConvNeXt, ViT and MLP-Mixer over cross-entropy training in 28 of 33 configurations, by 1.3 points on average, and is on par in accuracy with snapshot distillation, the closest competitor. On these families it is better than both in likelihood, calibration and robustness to corruptions and small adversarial perturbations. For an SGD-trained ResNet-18 it lowers accuracy but still improves robustness to small adversarial perturbations. Separating the reheated schedule from relearning shows that relearning helps every family and that the losses come from reheats, which undo most of this network's fit to its training set. How early the accuracy of a cross-entropy run saturates largely separates the AdamW families from ResNet-18 but does not predict the gain within them. With little data, relearning from the latest snapshot without dreams, as snapshot distillation does, compounds overconfidence in the AdamW families from episode to episode. The memory and the dreams each curb most of this drift, and together they remove it.

Figures & tables

Explore similar work

Oct 7, 2026

Lightweight and Versatile Learned Optimization by Recombination of Gradient History

This paper presents a lightweight and versatile learned optimizer that dynamically recombines gradient history, represented as averages over disjoint time spans. The optimizer reduces the prediction space to one scalar coefficient per gradient average, shared by multiple parameters. Progressively averaging older gradients minimizes memory cost of long history, while keeping their contributions independently accessible. A 37k-parameter network trained in 0.87 GPU-hours generalizes zero-shot to unseen tasks, lowering validation loss by 9.1% and 0.4% on BERT-Tiny and GPT-Tiny, and improving test accuracy over Adam by 3.5 %p on a Vision Transformer and by 2.7 %p on average across nine graph models, with FLOPs overhead as low as 0.3%.
Aug 26, 2026

Distance Is Not Enough: Forget-Retain Alignment Gap Predicts LLM Relearning Robustness

Machine unlearning aims to make a model forget specific data, yet unlearned LLMs often fail to stay unlearned: brief fine-tuning can revive removed knowledge. Existing robustness predictors rely on global weight-space displacement, but distance alone can be misleading when random or destructive updates collapse performance. We argue that relearning robustness depends on update structure: robust unlearning should affect forget-critical weights while sparing retain-critical ones. We introduce the Forget-Retain Alignment Gap (FRAG), a training-free predictor that scores an update's forget-retain alignment without running a relearning attack, and separates selective from dense updates more reliably than global distance. Building on the forget-critical, retain-sparing principle, Forget-Retain Pruning (FRP) improves relearning robustness. Our results suggest that weight selectivity better explains robustness than distance alone. Code is available at https://github.com/Yi1-Chen/FRAG.
Mar 7, 2026

ACD-U: Asymmetric co-teaching with machine unlearning for robust learning with noisy labels

Deep neural networks are prone to memorizing incorrect labels during training, which degrades their generalizability. Although recent methods have combined sample selection with semi-supervised learning (SSL) to exploit the memorization effect -- where networks learn from clean data before noisy data -- they cannot correct selection errors once a sample is misclassified. To overcome this, we propose asymmetric co-teaching with different architectures (ACD)-U, an asymmetric co-teaching framework that uses different model architectures and incorporates machine unlearning. ACD-U addresses this limitation through two core mechanisms. First, its asymmetric co-teaching pairs a contrastive language-image pretraining (CLIP)-pretrained vision Transformer with a convolutional neural network (CNN), leveraging their complementary learning behaviors: the pretrained model provides stable predictions, whereas the CNN adapts throughout training. This asymmetry, where the vision Transformer is trained only on clean samples and the CNN is trained through SSL, effectively mitigates confirmation bias. Second, selective unlearning enables post-hoc error correction by identifying incorrectly memorized samples through loss trajectory analysis and CLIP consistency checks, and then removing their influence via Kullback--Leibler divergence-based forgetting. This approach shifts the learning paradigm from passive error avoidance to active error correction. Experiments on synthetic and real-world noisy datasets, including CIFAR-10/100, CIFAR-N, WebVision, Clothing1M, and Red Mini-ImageNet, demonstrate state-of-the-art performance, particularly in high-noise regimes and under instance-dependent noise. The code is publicly available at https://github.com/meruemon/ACD-U.