Continual Learning & Catastrophic Forgetting
A Standardized Transfer-Validation Benchmark for Methods Validated Only on Small-Scale Image Classification
Generated automatically from the limitations stated in 152 papers (CVPR, ICML, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Nearly every training-time method in the modern literature—continual learning, distillation, unlearning, robustness, federated aggregation—is validated exclusively on CIFAR/ImageNet-subset image classification with ResNet-18/50-class backbones, and each paper explicitly defers scale, architecture, and modality transfer to future work. Because every paper defers this simultaneously, the field has no evidence base at all for whether any of these methods survive the shift to transformers, higher resolutions, dense tasks, or non-vision data; practitioners choosing between methods are choosing based on measurements made in a regime none of them deploy in. The barrier is not one method's limitation but a collective, unpriced evaluation debt: nobody can tell which small-scale conclusions are regime-robust and which are artifacts of ResNet-18 on 32×32 images.
Why it matters
For the first time, claims like "works at scale" or "architecture-agnostic" become testable and comparable across subfields, and method selection stops resting on 32×32 ResNet evidence alone.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Build a fixed "transfer grid" — a small, frozen set of evaluation points spanning (backbone: ResNet-18, ViT-Ti/B; scale: CIFAR-100, ImageNet-1k; task: classification, one dense task; modality: vision, text) — and port 5–8 representative open-source methods from different training-time families onto it, measuring exactly which ones degrade, fail to train, or require re-tuning at each grid point. Deliverable: a public leaderboard with per-cell reproduction instructions and compute budgets.
- 2
Run a systematic "regime sensitivity" study: for each method, quantify the gap between its reported small-scale headline result and its result under backbone swap and scale-up, holding hyperparameters fixed vs. re-tuned, to separate intrinsic fragility from mere tuning burden. Measured as a transfer-degradation score per method.
- 3
Distill the grid results into practitioner guidance: which classes of method (e.g., those depending on feature statistics vs. loss geometry) transfer predictably, tested by predicting held-out grid cells from observed degradation patterns.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
The grid's own coverage is finite, so it risks becoming one more small benchmark that methods overfit to, and compute costs may force most cells to use reduced resolutions — recreating the very regime the project exists to escape.
Sub-problems
Cross-Architecture and Cross-Task Generalisation of Neural Pruning and NAS Methods
A vast body of neural architecture search (NAS) and structural pruning algorithms is evaluated exclusively on standard convolutional backbones (e.g., ResNet, MobileNet) on CIFAR and ImageNet classification. Because architectural primitives (e.g., self-attention vs. spatial convolutions) and loss landscapes differ substantially across modalities and tasks, it is currently unknown whether these methods generalize, degrade, or completely fail when applied to Transformers, object detection backbones, or language tasks. Without systematic cross-paradigm evaluation, researchers and practitioners cannot determine whether published NAS/pruning performance gains reflect general compression principles or narrow overfitting to standard 2D-CNN benchmarks.
Benchmarking Dataset Distillation Across Image Resolutions, Dataset Scales, and Model Families
Dataset distillation methods are almost exclusively developed and evaluated on low-resolution toy benchmarks (CIFAR, Tiny-ImageNet at 32x32) using shallow 3-layer ConvNets. Consequently, it is unknown whether the relative performance of different distillation paradigms (e.g., gradient matching, trajectory matching, distribution matching) holds at high resolutions, on full ImageNet-1K scale, or when transferring to modern vision architectures like Vision Transformers. Practitioners cannot rely on synthetic datasets for real-world downstream training or architecture search because performance rankings established on toy regimes may be artifacts of shallow convolutional surrogates.
Cross-Architecture and Scale Robustness of Continual Learning Algorithms
The continual learning literature currently evaluates algorithms almost exclusively on small convolutional backbones (primarily ResNet-18) and downsampled vision benchmarks (CIFAR-100, Tiny-ImageNet). It is unknown whether canonical algorithmic rankings—such as the relative efficacy of experience replay, weight regularization, and parameter isolation—persist when transitioning to modern transformer backbones, full-scale image datasets, or non-vision modalities. Without systematic cross-architecture evaluation, continual learning research risks optimizing techniques for artifacts specific to low-capacity CNNs on small-scale vision splits rather than addressing catastrophic forgetting in modern foundational architectures.
A Systematic Generalization Benchmark for Dataset Distillation Across Dense Prediction and Non-Visual Modalities
Dataset distillation algorithms compress large training sets into compact synthetic samples, but current literature assesses them almost exclusively on image classification at low resolutions (≤128×128). Dense visual tasks such as object detection and semantic segmentation require fine-grained spatial and contextual preservation, while text and graph modalities require maintaining discrete structural semantics. Without empirical evaluation across these settings, practitioners cannot know whether current distillation objectives generalize to structured tasks or whether synthetic samples capture only global class-discriminative shortcuts.
Cross-Modality and Large-Scale Empirical Robustness of Federated Learning Algorithms
The vast majority of federated learning (FL) algorithms are evaluated exclusively on small vision datasets (such as CIFAR-10/100 and MNIST) using small CNN backbones. As a result, it is unknown whether established FL aggregation, regularization, and personalization techniques remain effective when applied to other modalities (natural language, tabular, audio) or modern large architectures (transformers, pre-trained backbones). Practitioners and researchers cannot reliably predict which algorithmic components fail or generalize beyond small-scale computer vision benchmarks.
Benchmarking Noisy and Partial Label Learning Across Model Architectures, Scale, and Modalities
The vast majority of algorithms for learning with noisy and partial labels are validated exclusively on low-resolution image benchmarks (CIFAR-10/100, TinyImageNet) using small convolutional backbones like ResNet-18. Consequently, practitioners have no reliable evidence on whether these loss corrections, sample-selection heuristics, or label-disambiguation schemes retain their relative performance on modern Transformer architectures, larger dataset scales, or non-vision modalities (NLP and tabular). Without cross-setting evaluation, methods tailored to small ResNets risk being brittle artifacts of low-capacity, small-resolution image regimes.
Systematic Evaluation of Adversarial Robustness and Transferability Across Architecture Families and Dataset Scales
Adversarial attack and defense mechanisms are predominantly validated on legacy convolutional neural networks (CNNs) and small 1,000-image ImageNet subsets. Because these methods are rarely evaluated on modern non-convolutional backbones such as Vision Transformers or across full-scale datasets and varying perturbation norms, their true cross-model efficacy remains unknown. Consequently, practitioners deploying defenses on modern vision architectures have no empirical guarantees on whether reported robustness gains transfer or cause unacceptable natural accuracy degradation.
Cross-Task and Cross-Modal Robustness Evaluation of Self-Supervised Representations
Core empirical findings and design choices in self-supervised and contrastive learning are almost exclusively established on ImageNet classification. It is unknown whether the standard inductive biases, augmentations, and objective functions that optimize classification transfer effectively to dense visual tasks (e.g., detection, segmentation) or non-vision modalities (e.g., audio, text). Without systematic cross-task and cross-modal evaluation, the community risks standardizing architectures and pre-training recipes that overfit to categorical image benchmarking while failing in broader operational settings.
Cross-Domain and Scale Generalization of Continual Learning Methods
Current continual learning techniques—such as rehearsal buffers, regularizers, and architecture-expansion methods—are almost exclusively evaluated on small, low-resolution image classification datasets (e.g., CIFAR-10/100, TinyImageNet) using small CNN backbones. Practitioners cannot determine whether these methods preserve performance when applied to dense vision tasks (detection, segmentation), language tasks, modern transformer architectures, or long-horizon task streams. As a result, deployment in real-world continuous adaptation pipelines remains blocked because algorithmic performance rankings derived from toy classification fail to transfer predictably to realistic settings.
Empirical Generalization of Adversarial and Backdoor Attacks Across Modern Vision Architectures and Scaled Benchmarks
The vast majority of published claims regarding backdoor attacks, defensive unlearning, and adversarial robustness rely exclusively on low-resolution datasets (e.g., CIFAR-10/100, GTSRB) evaluated on small CNN backbones (predominantly ResNet-18). It is currently unknown whether these attack success rates, perturbation budgets, and defensive mechanisms persist when applied to modern architectural paradigms (e.g., Vision Transformers, ConvNeXt) or larger-scale, higher-resolution tasks. As a result, security assurances and vulnerability assessments in real-world computer vision deployments remain ungrounded in empirical evidence.
Cross-Scale and Cross-Modal Robustness Benchmarking for Machine Unlearning
Current machine unlearning algorithms are evaluated almost exclusively on small vision datasets (e.g., CIFAR-10/100, TinyImageNet) using compact models (e.g., ResNet-18) at low removal ratios. Because unlearning dynamics depend heavily on parameter overcapacity and representation geometry, empirical efficacy on toy classification cannot predict behavior on realistic scales or other modalities. Practitioners cannot deploy unlearning algorithms to fulfill legal or privacy compliance on full-scale architectures because it remains unknown which methods maintain accuracy and forget guarantees outside of small-scale vision.
Cross-Paradigm and Scale Generalisation Benchmarks for Dataset Distillation
Dataset distillation methods are almost exclusively evaluated on small-scale datasets (CIFAR-10/100, Tiny-ImageNet) and homogeneous small ConvNet probe architectures. It is entirely unknown whether distilled synthetic data retains utility when transferred to Vision Transformers, modern hybrid backbones, or higher-resolution regimes with large capacity mismatches. Without systematic cross-architecture evaluation, distillation methods risk overfitting to the inductive biases of shallow ConvNets and failing in realistic downstream training pipelines.
Cross-Task Evaluation and Adaptation of Machine Unlearning Beyond Image Classification
Current machine unlearning algorithms are designed and validated almost exclusively on image classification benchmarks. It remains empirically unknown whether standard unlearning mechanisms—such as influence function approximations, gradient ascent, or Fisher-guided parameter scrubbing—generalise to structured vision tasks like object detection and semantic segmentation. Without systematic evaluation, practitioners cannot determine whether existing unlearning methods corrupt spatial representations, fail to erase localised object concepts, or cause catastrophic utility loss on dense prediction pipelines.
Cross-Architecture and Cross-Task Robustness Evaluation of Federated Learning Methods
Current federated learning methods for robustness and client heterogeneity are evaluated almost exclusively on standard convolutional backbones (primarily ResNet-18/50) across image classification datasets. As a consequence, it is unknown whether the mechanisms and empirical gains reported in the literature depend on CNN-specific properties (such as batch normalization statistics or localized feature representations) or if they hold across modern architectures like Vision Transformers and dense prediction tasks like segmentation and detection. Deploying existing FL techniques into production systems using non-CNN backbones or non-classification tasks carries unquantified risks of performance degradation and instability.
Systematic Cross-Architecture and High-Resolution Benchmark for Dataset Distillation
Dataset distillation methods are almost exclusively evaluated on tiny, low-resolution datasets (such as CIFAR-10 and MNIST) using identical 3-layer convolutional architectures for both synthesis and evaluation. Because existing studies do not evaluate performance when scaling to high-resolution images (such as ImageNet-scale) or transferring synthetic sets to modern deep backbones (such as ResNets and Vision Transformers), it is unknown whether current distillation objectives learn generalizable visual features or exploit inductive biases specific to shallow ConvNets. Without systematic empirical evaluation across diverse resolutions and model families, practitioners cannot determine whether synthetic datasets are viable for downstream training on standard deep learning pipelines.
Empirical Scaling and Generalisation Limits of Dataset Distillation Beyond Small Vision Benchmarks
Current dataset distillation algorithms are evaluated almost exclusively on low-resolution image classification datasets (CIFAR-10/100, Tiny ImageNet) using small ConvNets or shallow ResNets. Consequently, it is unknown whether synthesized datasets transfer effectively to large-scale vision backbones, modern generative models, or language models, despite broad motivational claims. Without cross-scale and cross-architecture evaluation, practitioners cannot determine whether observed distillation gains are algorithmic breakthroughs or artifacts of low-resolution, low-capacity regimes.
Empirical Robustness and Scalability Benchmark for Self-Supervised Learning Methods
Self-supervised learning algorithms are predominantly developed and evaluated on small backbones (ResNet-18) and downscaled subsets (ImageNet-100, STL-10, CIFAR), with gains often depending on setting-specific hyperparameter tuning. Because these methods have not been evaluated across larger backbones (such as Vision Transformers) and full-scale benchmarks (ImageNet-1k) under a unified protocol, it is unknown whether reported algorithmic advantages transfer or are artifacts of small-scale proxy setups. Without systematic multi-scale evaluation, researchers cannot determine which algorithmic mechanisms reliably generalize when scaled.
Evidence
Each paper's own statement of the limitation, verbatim.
- Two-Stage Unit Tying for Simplifying Differentiable Logic Gate NetworksICML 2026
Evaluated only on small convolutional logic networks on CIFAR-10 and MNIST (top accuracy 71.57%), so scaling to larger tasks or architectures is untested
- Social Hippocampus Memory LearningICML 2026
Experiments restricted to two image classification benchmarks with CNN-based heterogeneous models; no text, tabular, or transformer settings tested.
- Smaller but Better: Plasticity-Preserving Continual Learning for Embedded AITheWebConference 2026
Evaluation is restricted to small CNNs (ResNet-18/34/50) on CIFAR-10/100 and TinyImageNet resized to 32x32; no results on transformers, other modalities, or real deployment task streams
- Multimodal Distribution Matching for Vision-Language Dataset DistillationCVPR 2026
Assumes access to pretrained image and text encoders; performance is bounded by these pretrained models.
- Mitigating The Distribution Shift of Diffusion-based Dataset DistillationCVPR 2026
Evaluated only on image classification with ResNet/ConvNet probes at small IPC (10-100); cross-architecture generalization is argued rather than experimentally tested
- Learnability and Privacy Vulnerability are Entangled in a Few Critical WeightsICLR 2026
Empirical validation is restricted to small-resolution (32x32) image classification tasks and small architectures (ResNet18, ViT with 4x4 patches).
- IDER: IDEMPOTENT EXPERIENCE REPLAY FOR RELIABLE CONTINUAL LEARNINGICLR 2026
Evaluation is limited to small-scale image benchmarks (CIFAR-10/100, Tiny-ImageNet) with ResNet-18; no evidence on large-scale or non-vision domains
- Grounding and Enhancing Informativeness and Utility in Dataset DistillationICLR 2026
Evaluated only on image classification benchmarks; extension to other data modalities and real-world complex datasets is left to future work
Show all 152 papers
- Grounding Functional Similarity by Invariance-Aware Model StitchingICML 2026
Experiments limited to image classification models; other tasks and modalities unexplored.
- FedPAT: Federated Test-Time Adaptation via Prototype Affinity TopologyICML 2026
Evaluated only on image corruption benchmarks (CIFAR-10-C, CIFAR-100-C, Tiny-ImageNet-C) with ResNet18 backbone in main experiments; generalization to other modalities or architectures is only partially explored (appendix ResNet50/ViT-B/16)
- FedCART: Tackling Long-Tailed Distributions in Federated Adversarial Training via Classifier RefinementCVPR 2026
Evaluation limited to small-scale image datasets (CIFAR-10/100, FMNIST, SVHN) with only 5–20 clients and ResNet-18 or lightweight CNN backbones
- FaLW: A Forgetting-aware Loss Reweighting for Long-tailed UnlearningICLR 2026
Evaluated only on image classification benchmarks (CIFAR-10, CIFAR-100, Tiny-ImageNet) using standard CNN architectures (ResNet-18, VGG-16).
- Exploring Nonlinear Pathway in Parameter Space for Machine UnlearningICML 2026
Evaluated only on image classification benchmarks (CIFAR-10, ImageNet-100) using CNN and ViT architectures.
- Diffusion Models as Dataset Distillation PriorsICLR 2026
Evaluated only on image classification datasets (ImageNet-1K and subsets); extension to language, video, and multimodal data is left to future work
- Differentiable JPEG-based Input Perturbation for Knowledge Distillation Amplification via Conditional Mutual Information MaximizationICLR 2026
Evaluation is limited to image classification, leaving dense visual tasks (e.g., object detection, segmentation) untested
- Dataset Distillation by Influence MatchingCVPR 2026
Classification experiments use small ConvNet backbones (cross-architecture tests on CIFAR-100 only); largest image benchmark is Tiny-ImageNet
- Beyond Myopic Alignment: Lookahead Optimization for Online Class-Incremental LearningCVPR 2026
Evaluation is limited to image classification benchmarks with ResNet-18 backbones; no results on other modalities or larger architectures
- Beyond Buffer Limits: Energy-Based Data Reassembly for Continual LearningICML 2026
Evaluated only on small-resolution image classification (32×32–224×224, ≤200 classes); applicability to dense tasks (segmentation, detection) or text is untested
- Asymmetric Synthetic Data Update for Domain Incremental Dataset DistillationICLR 2026
Empirical validation is limited to low-resolution (32x32) images and shallow 3-layer ConvNet architectures.
- Adaptive Width Neural NetworksICLR 2026
Formulated and evaluated strictly for Multi-Layer Perceptrons and MLP sub-modules, leaving direct extensions to convolutional filters or attention heads to future work.
- Versatile Transferable Unlearnable Example GeneratorNeurIPS 2025
Evaluation is limited to small-scale classification benchmarks (CIFAR-10/100, SVHN, PACS) with ResNet-18 as surrogate and target model; unlearnability on large-scale or diverse architectures is not demonstrated in the provided text
- VL2Lite: Task-Specific Knowledge Distillation from Large Vision-Language Models to Lightweight NetworksCVPR 2025
Evaluation is restricted to image classification, leaving performance on dense prediction tasks (segmentation, detection) unverified.
- Unveiling AI’s Blind Spots: An Oracle for In-Domain, Out-of-Domain, and Adversarial ErrorsICML 2025
Evaluated only on image classification (CIFAR-10/100, ImageNet-1K); generalization to other tasks like detection or translation is untested
- Unlocking Dataset Distillation with Diffusion ModelsNeurIPS 2025
Evaluation limited to 10-class ImageNet subsets and CIFAR-10; ImageNet-1K scale is untested
- Unlearn and Burn: Adversarial Machine Unlearning Requests Destroy Model AccuracyICLR 2025
Evaluated primarily on image classification tasks rather than language or tabular data.
- Two is Better than One: Efficient Ensemble Defense for Robust and Compact ModelsCVPR 2025
Evaluations are limited to small-scale image classification (CIFAR-10, SVHN) with ResNet18/VGG16; no larger datasets or architectures are tested
- Towards flexible perception with visual memoryICML 2025
Only image classification is considered; detection, segmentation, and generative tasks are untested
- Towards Stable and Storage-efficient Dataset Distillation: Matching Convexified TrajectoryCVPR 2025
Evaluated only on small 32x32/64x64 image datasets with a small ConvNet surrogate; claims of benefit for large models (e.g., LLMs) are speculative and untested
- Task-Agnostic Guided Feature Expansion for Class-Incremental LearningCVPR 2025
Evaluation limited to image classification on CIFAR100/ImageNet with ResNet18 backbones; transfer to other domains or architectures not demonstrated
- Single Teacher, Multiple Perspectives: Teacher Knowledge Augmentation for Enhanced Knowledge DistillationICLR 2025
Only validated on image classification benchmarks (CIFAR100, ImageNet, TinyImageNet, STL10); applicability to other modalities is untested
- Simple Distillation for One-Step Diffusion ModelsNeurIPS 2025
Evaluated only on 32x32 CIFAR-10 and 64x64 ImageNet; no text-to-image or large-scale results, which the authors leave to future work
- SelKD: Selective Knowledge Distillation via Optimal Transport PerspectiveICLR 2025
Evaluated only on small-scale classification benchmarks (CIFAR-10, CIFAR-100, Tiny ImageNet) with ResNet teachers; no foundation-model-scale experiments despite that motivation
- Rethinking Epistemic and Aleatoric Uncertainty for Active Open-Set Annotation: An Energy-Based ApproachCVPR 2025
Evaluated only on image classification benchmarks (CIFAR-10/100, Tiny-ImageNet) with ResNet-18; hyperparameters (margins m_kno=-25, m_unk=-7, ω_e=0.01, K=250) may not transfer to other domains
- Resource-Constrained Federated Continual Learning: What Does Matter?NeurIPS 2025
Datasets are image classification of modest scale (CIFAR, Tiny-ImageNet, Office variants, Digit-10); no language, detection, or very large-scale tasks are tested
- Projection Head is Secretly an Information BottleneckICLR 2025
Experiments are limited to ResNet-18 backbones on CIFAR-10/100 and ImageNet-100 with only SimCLR and Barlow Twins; no large-scale (ImageNet-1k) or modern architecture results
- Probabilistic Group Mask Guided Discrete Optimization for Incremental LearningICML 2025
Tested only on small CNN benchmarks (AlexNet on Split CIFAR-100, LeNet on Superclass); no evidence on large models or natural language tasks
- Phase and Amplitude-aware Prompting for Enhancing Adversarial RobustnessICML 2025
Reduces natural accuracy on naturally pre-trained models (e.g., 94.83% to 87.12% on CIFAR-10/ResNet18).
No ImageNet-scale evaluation due to computational constraints.
- Perceptual Inductive Bias Is What You Need Before Contrastive LearningCVPR 2025
Only ResNet18 on ImageNet-100 and STL-10 due to compute constraints; scalability to larger models and datasets is unverified
- One-Step Offline Distillation of Diffusion-based Models via Koopman ModelingNeurIPS 2025
Fails to scale experiments beyond CIFAR-10/FFHQ/AFHQ: ImageNet and GET baselines on higher-resolution data were excluded as computationally infeasible for the authors, so large-scale claims are untested
- On the Convergence of Continual Federated Learning Using Incrementally Aggregated GradientsAISTATS 2025
Experiments limited to task-incremental image classification benchmarks (ResNet-18/50 on CIFAR/TinyImageNet splits); class-incremental results deferred to supplementary
- NoT: Federated Unlearning via Weight NegationCVPR 2025
Evaluated only on vision classification with small datasets (CIFAR-10/100, Caltech-101); untested on large-scale data like ImageNet.
- Measuring Representational Shifts in Continual Learning: A Linear Transformation PerspectiveICML 2025
Experiments are limited to image classification with a small modified ResNet (8 channels, 9 blocks) on Split-CIFAR100 and downsampled ImageNet1K with N=50 tasks; no NLP, other architectures beyond brief ViT checks, or other continual learning setups are tested
- Leveraging Spatial Invariance to Boost Adversarial TransferabilityICCV 2025
Evaluated only on the 1,000-image ImageNet-compatible dataset with L∞ ε=16, T=10; no other datasets, norms, or image resolutions are tested
- Learning without Isolation: Pathway Protection for Continual LearningICML 2025
Not validated on large models (e.g., Transformers/LLMs); all experiments use ResNet18/ResNet32 on CIFAR-100 and Tiny-ImageNet
- Learnable Logit Adjustment for Imbalanced Semi-Supervised Learning under Class Distribution MismatchICCV 2025
Evaluated only on image classification benchmarks (CIFAR-10/100-LT, STL-10-LT, Small-ImageNet-127); applicability to tabular or NLP data is left as future work
- Improving Adversarial Transferability on Vision Transformers via Forward Propagation RefinementCVPR 2025
Evaluated only on ImageNet (1000 ILSVRC2012 validation images) and classification models; effectiveness on other tasks or smaller image sets is untested
- GraTeD-MLP: Efficient Node Classification via Graph Transformer Distillation to MLPLearning on Graphs 2025
Effectiveness is only demonstrated for node classification, with unknown performance on edge or graph-level tasks.
- Fuse Before Transfer: Knowledge Fusion for Heterogeneous DistillationICCV 2025
Validated only on image classification (CIFAR-100, ImageNet-1K); detection and NLP are untested
- Ferret: An Efficient Online Continual Learning Framework under Varying Memory ConstraintsCVPR 2025
Evaluated only on small-to-medium image-classification models (MLP, MNISTNet, ConvNet, ResNet-18, MobileNet); scaling behavior for large models or non-vision tasks is untested
- FedWSQ: Efficient Federated Learning with Weight Standardization and Distribution-Aware Non-Uniform QuantizationICCV 2025
Experiments are image classification only (CIFAR-10/100, Tiny-ImageNet with ResNet-18 and a few small backbones, 100 clients, 5% participation); no language or large-model settings are evaluated
- Enhancing Visual Prompting through Expanded Transformation Space and Overfitting MitigationNeurIPS 2025
Evaluated only on image classification; no other task types (detection, segmentation) are tested
- Enhancing Online Continual Learning with Plug-and-Play State Space Model and Class-Conditional Mixture of DiscretizationCVPR 2025
Evaluated only on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18 without pretraining; no evidence on larger datasets, pretrained backbones, or non-image modalities
- Enhancing Adversarial Robustness with Conformal Prediction: A Framework for Guaranteed Model ReliabilityICML 2025
Experimental scope is limited to ResNet34/ResNet50 on CIFAR-10, CIFAR-100, and mini-ImageNet with L-infinity budget 8/255; authors acknowledge the setup is relatively limited and plan other architectures.
- Dynamic Pseudo Labeling via Gradient Cutting for High-Low Entropy ExplorationCVPR 2025
Evaluated only on small image classification benchmarks (CIFAR, SVHN, STL-10) with Wide-ResNet encoders; generalization to other architectures or tasks untested
- Dataset Distillation with Neural Characteristic Function: A Minmax PerspectiveCVPR 2025
Evaluation is limited to image classification benchmarks up to 128×128 ImageNet subsets; no evidence on other modalities, larger-scale full ImageNet, or downstream tasks at scale.
- Dataset Distillation as Data Compression: A Rate-Utility PerspectiveICCV 2025
Evaluated only on image classification at ≤128x128 resolution; extension to video, text, and graphs requires new decoders and entropy models and is untested
- Data-free Universal Adversarial Perturbation with Pseudo-semantic PriorCVPR 2025
Evaluated only on CNN classifiers on ImageNet; no evidence on transformers, other datasets, or other tasks
- Curriculum Coarse-to-Fine Selection for High-IPC Dataset DistillationCVPR 2025
Evaluated only on small image datasets (CIFAR-10/100, Tiny-ImageNet) at 5-30% compression ratios; very low IPC or larger datasets are untested
- Confusion-Driven Self-Supervised Progressively Weighted Ensemble Learning for Non-Exemplar Class Incremental LearningNeurIPS 2025
Evaluated only on ResNet18 at moderate scale (CIFAR100, TinyImageNet, 100-class ImageNet-Subset); large-scale or complex scenarios left to future work
- CAN: Leveraging Clients As Navigators for Generative Replay in Federated Continual LearningICML 2025
Evaluated only on image classification benchmarks (CIFAR100, TinyImageNet, ImageNet100) with class-incremental splits; no evidence on other modalities or task types
- Boosting Adversarial Transferability via Residual Perturbation AttackICCV 2025
Evaluation relies on only 1,000 ImageNet validation images and a limited set of surrogate/target models, following prior work conventions
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and ApproximationICLR 2025
Evaluation limited to relatively small image classification datasets (CIFAR100, TinyImageNet, ImageNet at reduced 64x64 resolution) and small storage budgets; behavior at large scale or with high-resolution data is not demonstrated in the main results
- Beyond Low-rank Decomposition: A Shortcut Approach for Efficient On-Device LearningICML 2025
Evaluated only on small CNNs (MCUNet, ResNet-18/34, MobileNetV2); extension to larger models such as LLMs is left as future work
- A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced DatasetsCVPR 2025
Evaluation limited to binary image classification; performance on multi-class or non-image modalities is untested
- What is Dataset Distillation Learning?ICML 2024
Analyses are primarily on CIFAR-10 with a small 3-layer ConvNet and 10 images per class, with additional datasets deferred to the appendix
- Unraveling the Key Components of OOD Generalization via DiversificationICLR 2024
Evidence comes mainly from image classification benchmarks; generalization of conclusions to other data types is asserted but not tested
- Ungeneralizable ExamplesCVPR 2024
Framework only designed for image classification; extension to other tasks is left as future work
- Understanding and Mitigating the Label Noise in Pre-training on Downstream TasksICLR 2024
Analytical experiments use only ResNet-50 and relatively small-scale pre-training (ImageNet-1K, YFCC15M, ImageNet-21K) due to limited compute; results on larger foundation models are unverified
- Towards Scalable and Versatile Weight Space LearningICML 2024
Experiments restricted to computer vision classification tasks only
- Towards Realistic Model Selection for Semi-supervised LearningICML 2024
Empirically validated only on small image benchmarks (CIFAR-10/100, Semi-Aves) with WideResNet-28-2; generalization to other architectures or modalities is not shown
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce EdgeICML 2024
Evaluated only on CNN architectures for vision classification; Transformers, RNNs, and other tasks (segmentation, audio) are untested
- Small Scale Data-Free Knowledge DistillationCVPR 2024
Evaluated only on CIFAR-10/100 classification and a single small-scale NYUv2 segmentation task; generalization to large-scale datasets or other architectures is untested.
- Self-Supervised Adversarial Training via Diverse Augmented Queries and Self-Supervised Double PerturbationNeurIPS 2024
Evaluated only on CIFAR-10, CIFAR-100, and STL-10 with ResNet backbones; no large-scale or higher-resolution validation
- SUMix: Mixup with Semantic and Uncertain InformationECCV 2024
Validated only on image classification; not shown for detection, segmentation, or semi-supervised settings
- Revisit the Essence of Distilling Knowledge through CalibrationICML 2024
Experiments are limited to CIFAR-10/100 and Tiny-ImageNet with ResNet/WRN/MobileNet families; no large-scale or real capacity-mismatch regime (e.g., vision transformers teaching tiny students) is tested
- Regularizing with Pseudo-Negatives for Continual Self-Supervised LearningICML 2024
Only CNN (ResNet) architectures and the vision domain evaluated; ViT and NLP settings deferred to future work
- Preventing Dimensional Collapse in Self-Supervised Learning via Orthogonality RegularizationNeurIPS 2024
Large-scale results limited to BYOL with ResNet50 at 100-200 epochs and 128-256 batch size; reported gains rely on per-setting gamma tuning on a validation set
- Orchestrate Latent Expertise: Advancing Online Continual Learning with Multi-Level Supervision and Reverse Self-DistillationCVPR 2024
Evaluated only with ResNet18 backbone on CIFAR-100 and Tiny-ImageNet class-incremental splits; generalization to ViT or other backbones is left to future work
- Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster MemoryCVPR 2024
Experiments are limited to 32×32 and 64×64 images on relatively small datasets; no evaluation at high-resolution scale or on very long streams
- Meta Continual Learning Revisited: Implicitly Enhancing Online Hessian Approximation via Variance ReductionICLR 2024
Results are limited to small-image classification benchmarks (CIFAR10, CIFAR100, TinyImageNet) with small networks; no evidence on other domains or task types.
- MaxQ: Multi-Axis Query for N:M Sparsity NetworkCVPR 2024
Evaluated only on CNNs for computer vision tasks (ResNet50 and detection/segmentation backbones); no evidence for transformers, NLP, or other architectures
- Make Continual Learning Stronger via C-FlatNeurIPS 2024
Evaluated only on class-incremental image classification benchmarks (CIFAR-100, ImageNet-100, Tiny-ImageNet) with ResNet-32/18; no NLP or other modality evidence
- Long-Tail Class Incremental Learning via Independent Sub-prototype ConstructionCVPR 2024
Only evaluated on 100-class datasets (CIFAR100, ImageNet-Subset) with ResNet18 and small task counts (5 or 10 tasks); scalability to many-task or larger-scale settings is untested
- Logit Standardization in Knowledge DistillationCVPR 2024
Evaluation limited to classification (CIFAR-100, ImageNet, plus supplements on CUB200/COCO); no dense prediction or non-vision tasks in the main results
- Learning to Unlearn for Robust Machine UnlearningECCV 2024
Evaluated only on image classification (CIFAR-10, Tiny-ImageNet) with ResNet-18/ViT; no evidence on text, large-scale models, or non-vision domains
- Improving Robustness to Model Inversion Attacks via Sparse Coding ArchitecturesECCV 2024
Experiments are on linear-network targets; generalization to deeper, application-specific architectures (e.g., ResNets, Transformers) is not demonstrated
- HYDRA-FL: Hybrid Knowledge Distillation for Robust and Accurate Federated LearningNeurIPS 2024
Evaluated only on small image datasets (MNIST, CIFAR-10, CIFAR-100) with small CNNs; no language, multimodal, or larger-scale models
- Gradual Divergence for Seamless Adaptation: A Novel Domain Incremental Learning MethodICML 2024
Evaluated only on two datasets (DN4IL, iCIFAR-20) with ResNet-18; no results on larger backbones or class-incremental settings
- FedSOL: Stabilized Orthogonal Learning with Proximal Restrictions in Federated LearningCVPR 2024
Evaluated only on small image classification tasks (MNIST/CIFAR/SVHN/CINIC/MedMNIST) with a small CNN; behavior on large-scale or non-image tasks is unverified
- FedMef: Towards Memory-efficient Federated Dynamic PruningCVPR 2024
Only evaluated on small-scale image classification datasets (CIFAR-10, CINIC-10, TinyImageNet) with ResNet18 and MobileNetV2; no large-scale or non-vision tasks
- FedBAT: Communication-Efficient Federated Learning via Learnable BinarizationICML 2024
Experiments limited to small-scale image datasets (FMNIST, SVHN, CIFAR-10/100) with CNN and ResNet-10; no evaluation on large models, language tasks, or realistic system conditions
- Distribution-Aware Robust Learning from Long-Tailed Data with Noisy LabelsECCV 2024
Evaluated only on CIFAR-scale (32x32) datasets and Red mini-ImageNet with PreAct ResNet-18/ResNet-18 backbones; no large-scale ImageNet-class experiments
- Deep Feature Surgery: Towards Accurate and Efficient Multi-Exit NetworksECCV 2024
Evaluation is limited to image classification with CNN backbones (VGG, ResNet18, MSDNet) on Cifar100 and ImageNet; applicability to other architectures or tasks is untested
- Dataset Distillation by Automatic Training TrajectoriesECCV 2024
Evaluated only on image classification benchmarks up to 128x128 ImageNet subsets with small ConvNets; no transformer or larger-scale results
- Data Poisoning based Backdoor Attacks to Contrastive LearningCVPR 2024
Experiments limited by compute to a 100-class ImageNet subset and ResNet-18 default; reference images are assumed drawn from the target downstream task's distribution
- DEPrune: Depth-wise Separable Convolution Pruning for Maximizing GPU ParallelismNeurIPS 2024
Only evaluated on image classification CNNs (MobileNet-V2/V3, EfficientNet-B0) with DSConv; applicability to other architectures or tasks is untested
- Correlation-Induced Label Prior for Semi-Supervised Multi-Label LearningICML 2024
Evaluated only on image classification benchmarks (VOC, COCO, NUS-WIDE) with ResNet-50 features; applicability to other modalities or label structures is untested
- Continual Learning in the Frequency DomainNeurIPS 2024
Evaluated only on small-image benchmarks (32x32 CIFAR-10 and Tiny ImageNet splits), so applicability to larger-resolution or non-image domains is untested
- Bidirectional Uncertainty-Based Active Learning for Open-Set AnnotationECCV 2024
Validated only on image classification (CIFAR-10, CIFAR-100, Tiny-Imagenet) with ResNet18; performance on other modalities or large-scale datasets is not shown
- Auto-Train-Once: Controller Network Guided Automatic Network Pruning from ScratchCVPR 2024
Evaluated only on CNN image classification (ResNet/MobileNetV2 on CIFAR/ImageNet); no transformers, detection, or other tasks
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned ModelsCVPR 2024
Evaluated primarily on convolutional neural networks and standard image classification benchmarks; applicability to non-CNN architectures (e.g., Vision Transformers) or other domains is unverified.
- Approximate Bayesian Class-Conditional Models under Continuous Representation ShiftAISTATS 2024
Only evaluated on small-scale image datasets (CIFAR-10, CIFAR-100, MiniImageNet) with a reduced ResNet18; no large-scale or non-vision evaluation
- Aligned Objective for Soft-Pseudo-Label Generation in Supervised LearningICML 2024
Only validated on small image benchmarks (CIFAR-10/100, TinyImageNet, 32x32/64x64 resolution with ResNet-20/44/18 backbones); scalability to larger models or datasets is not shown
- Adversarial Feature Map Pruning for BackdoorICLR 2024
Evaluated only on ResNet18/PreActResNet18 image classifiers on CIFAR-10/100 and GTSRB; applicability to other architectures and modalities is untested
- A Label is Worth A Thousand Images in Dataset DistillationNeurIPS 2024
Only image classification tasks evaluated; generalization to other modalities is untested
- Wasserstein Expansible Variational Autoencoder for Discriminative and Generative Continual LearningICCV 2023
Evaluation is limited to small image benchmarks (MNIST, Fashion, CIFAR, MiniImageNet splits); scalability to long, large-scale streams or non-image modalities is not demonstrated
- UMD: Unsupervised Model Detection for X2X Backdoor AttacksICML 2023
Evaluation restricted to image classification (ResNet-18 / GTSRB CNN) with relatively small datasets (CIFAR-10, GTSRB, Imagenette); no evaluation on large-scale or non-image domains
- Trainable Weight Averaging: Efficient Training by Optimizing Historical SolutionsICLR 2023
Evaluated only on image classification (CIFAR-10/100, ImageNet) with CNN architectures; no evidence on other tasks (NLP, detection) or transformer architectures
- Towards Stable and Efficient Adversarial Training against $l_1$ Bounded Adversarial AttacksICML 2023
Robust accuracy results shown only for small datasets (CIFAR10/100, ImageNet100 subset); full ImageNet evaluation not reported
- Towards Addressing Label Skews in One-Shot Federated LearningICLR 2023
Evaluated only on small image datasets (MNIST, FMNIST, SVHN, CIFAR-10) with a 5-layer CNN; no large-scale or non-image results
- The Curious Case of Benign MemorizationICLR 2023
Experiments are limited to small vision benchmarks (CIFAR10, CIFAR100, TinyImageNet) with convolutional encoders; generalization to other architectures, modalities, or large-scale data is untested
- Supervision Complexity and its Role in Knowledge DistillationICLR 2023
Experiments are restricted to small-scale image classification (CIFAR-10/100, Tiny ImageNet); large-scale or non-vision settings are untested
- Structural Alignment for Network Pruning through Partial RegularizationICCV 2023
Evaluated only on classification CNNs (ResNets, MobileNet-V2) on CIFAR-10/ImageNet; no evidence for other tasks or architectures.
- StableFDG: Style and Attention Based Learning for Federated Domain GeneralizationNeurIPS 2023
Evaluated only with ResNet-18/50 backbones on image classification benchmarks; CNN style statistics assumptions may not transfer to other architectures or tasks (segmentation/detection left to future work)
- Simulated Annealing in Early Layers Leads to Better GeneralizationCVPR 2023
Results are limited to ResNet50 on Tiny-ImageNet (64x64 images); no evidence on full ImageNet or larger/other architectures
- RoPAWS: Robust Semi-supervised Representation Learning from Uncurated DataICLR 2023
Evaluated only on image classification with ResNet-50 / WRN-28-2; extension to other modalities is not shown
- Revisiting Visual Model Robustness: A Frequency Long-Tailed Distribution ViewNeurIPS 2023
Findings and method are validated only on image classification; extension to other vision tasks is left as future work
- Revisiting Adversarial Robustness Distillation from the Perspective of Robust FairnessNeurIPS 2023
Evaluation limited to small-scale benchmarks (CIFAR-10 shown in main text; CIFAR-100, SVHN, Tiny-ImageNet relegated to appendix)
- Prototypical Mixing and Retrieval-Based Refinement for Label Noise-Resistant Image RetrievalICCV 2023
Evaluated only with ResNet18 backbones on relatively small datasets (up to 60k images); scalability to web-scale retrieval databases is untested
- Prototype-Sample Relation Distillation: Towards Replay-Free Continual LearningICML 2023
Evaluation restricted to vision benchmarks (CIFAR-100, MiniImageNet, ImageNet32); no evidence on other modalities or online/streaming settings
- Practical Network Acceleration With Tiny SetsCVPR 2023
Demonstrated only on classification with residual-block-structured CNNs (ResNet, MobileNetV2); Transformers and other tasks (detection, segmentation) are explicitly left as future work.
- Personalized Semantics Excitation for Federated Image ClassificationICCV 2023
Experiments use small CNNs and digit/CIFAR-scale datasets (up to 200 clients); no evaluation on large-scale data, deep networks, or heterogenous client compute.
- Personalized Federated Learning with Feature Alignment and Classifier CollaborationICLR 2023
Evaluated only on classification tasks with CNNs on four image datasets; extension to other tasks/model families and dynamic/decentralized settings is left to future work
- On the Stability-Plasticity Dilemma of Class-Incremental LearningCVPR 2023
Analysis is limited to ResNet-18 on ImageNet-1K with 20 exemplars per class; no transformer architectures or other datasets in the main results
- ORC: Network Group-based Knowledge Distillation using Online Role ChangeICCV 2023
Evaluated only on image classification (CIFAR-10/100, ImageNet) with CNN backbones; applicability to other tasks or architectures is untested.
- Masked Autoencoders Enable Efficient Knowledge DistillersCVPR 2023
Evaluated mainly on ImageNet classification; downstream validation limited to a single COCO Mask R-CNN setup
- Make Landscape Flatter in Differentially Private Federated LearningCVPR 2023
Experiments limited to small/moderate image datasets (EMNIST, CIFAR-10/100) with CNN/ResNet-18 and 500 clients; large-scale models or text tasks untested
- MIM4DD: Mutual Information Maximization for Dataset DistillationNeurIPS 2023
Evaluated only on small 28x28/32x32 benchmarks (MNIST, SVHN, CIFAR10/100) with 3-layer ConvNets; no ImageNet-scale or cross-architecture generalization tests
- Learning From Biased Soft LabelsNeurIPS 2023
Results are shown only on CIFAR-10, CIFAR-100, and Tiny ImageNet; effectiveness on larger-scale datasets or other architectures beyond WideResNet variants is not demonstrated
- Label-Only Model Inversion Attacks via Knowledge TransferNeurIPS 2023
Only evaluated on face recognition targets with 3 architectures (VGG16, IR152, FaceNet64, plus Resnet-152 at 128×128); other data types such as healthcare data are untested
- Improved Visual Fine-tuning with Natural Language SupervisionICCV 2023
Evaluated only on classification; extension to other vision tasks is left as future work
- IPMix: Label-Preserving Data Augmentation Method for Training Robust ClassifiersNeurIPS 2023
Evaluated only on CIFAR/ImageNet-scale classification with specific architectures (WRN, ResNeXt, ResNet-18/50); no transformer results reported
- GPFL: Simultaneously Learning Global and Personalized Feature Information for Personalized Federated LearningICCV 2023
Evaluated with small backbones (4-layer CNN, fastText, 3-layer MLP, HAR-CNN) and ResNet-18; no experiments with large pre-trained models or very deep networks
- From Knowledge Distillation to Self-Knowledge Distillation: A Unified Approach with Normalized Loss and Customized Soft LabelsICCV 2023
Evaluation is limited to classification (plus one COCO detection transfer); no results on other tasks or very large-scale settings beyond ImageNet
- Frequency Domain-Based Dataset DistillationNeurIPS 2023
Evaluation is limited to small-scale classification (CIFAR, SVHN, ImageNet subsets, 3D MNIST); no full-scale ImageNet or non-classification tasks are reported
- Federated Learning with Bilateral Curation for Partially Class-Disjoint DataNeurIPS 2023
Evaluated only with ResNet18 on image classification benchmarks and ISIC; performance on other modalities or architectures is untested
- Fast Federated Machine Unlearning with Nonlinear Functional TheoryICML 2023
Evaluation is limited to small image classification datasets (Fashion-MNIST, CIFAR-10, SVHN) with small models (CNN, LeNet, ResNet-18) at removal ratios ≤20%; scalability to large models and high removal rates is untested
- Eliminating Domain Bias for Federated Learning in Representation SpaceNeurIPS 2023
Evaluated only on small-scale classification (4-layer CNN, ResNet-18, fastText; FMNIST/Cifar100/Tiny-ImageNet/AG News, 20-100 clients) with full client participation by default; untested on large models or cross-device settings
- Downstream-agnostic Adversarial ExamplesICCV 2023
Evaluated only with ResNet18 backbones on four image classification/retrieval datasets; no evaluation on detection/segmentation or larger architectures
- Disposable Transfer Learning for Selective Source Task UnlearningICCV 2023
Experiments confined to small-to-medium-scale datasets (CIFAR, STL-10, SVHN, TinyImageNet) and ResNet-18, so effectiveness on realistic-scale models and data is unverified
- Dense Network Expansion for Class Incremental LearningCVPR 2023
Evaluated only on CIFAR100/ImageNet100 with a small 6-layer ViT and 2000-example buffer; no results on longer task sequences or larger backbones
- DeepMAD: Mathematical Architecture Design for Deep Convolutional Neural NetworkCVPR 2023
Restricted to conventional convolutional building blocks; generalization to attention/transformer blocks is left to future work
- Cumulative Spatial Knowledge Distillation for Vision TransformersICCV 2023
Evaluated only on classification-style tasks (ImageNet-1k and classification transfer benchmarks); no dense prediction tasks tested
- Contrastive Continuity on Augmentation Stability Rehearsal for Continual Self-Supervised LearningICCV 2023
Evaluation is limited to small low-resolution benchmarks (CIFAR-10/100, Tiny-ImageNet) with small buffers (200-256) and ResNet-18; no large-scale or long task-sequence results are reported
- Computationally Budgeted Continual Learning: What Does Matter?CVPR 2023
Experiments use a single architecture (ResNet50) and two image classification datasets (ImageNet2K, CGLM); generality to other modalities or architectures is not tested.
- Boosting Adversarial Transferability via Gradient Relevance AttackICCV 2023
Evaluated only on a 1,000-image subset of ILSVRC 2012 with ε=16, T=10 ImageNet classifiers; no other datasets or tasks are tested
- Blurred-Dilated Method for Adversarial AttacksNeurIPS 2023
Applies only to CNN models with convolutional and pooling layers; extension to Vision Transformers is left to future work.
- Bayesian Optimization Meets Self-DistillationICCV 2023
Evaluation is confined to small 32x32-image settings (CIFAR/Tiny-ImageNet resized to 32x32) and relatively small architectures (AlexNet, VGG, ResNet-20/56, ResNet-34, 13-CNN); no ImageNet-scale or large-model results
- Automatic Attention Pruning: Improving and Automating Model Pruning using AttentionsAISTATS 2023
Evaluated only on CNN image-classification architectures (ResNet, VGG, MobileNet, ShuffleNet, LeNet) and image datasets; no evidence for transformers, detection, or NLP tasks
- An Efficient Dataset Condensation Plugin and Its Application to Continual LearningNeurIPS 2023
Evaluation is limited to small benchmarks (MNIST, CIFAR10/100, TinyImageNet) with ConvNet-style training; no ImageNet-scale or high-resolution experiments
- Powering One-shot Topological NAS with Stabilized Share-parameter ProxyECCV 2020
Restricted to 2D CNN classification and detection architectures, leaving other modalities and transformer architectures untested.
- SecureNets: Secure Inference of Deep Neural Networks on an Untrusted CloudACML 2018
Evaluation is limited to shallow 4-5 layer networks, leaving performance on modern very deep architectures unproven.
Nearest existing work
- Efficiently Robustify Pre-Trained ModelsICCV 2023
- Select and Distill: Selective Dual-Teacher Knowledge Transfer for Continual Learning on Vision-Language ModelsECCV 2024
- Catastrophic Forgetting Meets Negative Transfer: Batch Spectral Shrinkage for Safe Transfer LearningNeurIPS 2019
- Adversarially robust transfer learningICLR 2020
- Continual Normalization: Rethinking Batch Normalization for Online Continual LearningICLR 2022
- Batch Model Consolidation: A Multi-Task Model Consolidation FrameworkCVPR 2023
- DKT: Diverse Knowledge Transfer Transformer for Class Incremental LearningCVPR 2023
- Transfer without ForgettingECCV 2022
- On the Approximation Risk of Few-Shot Class-Incremental LearningECCV 2024
- Overcoming Catastrophic Forgetting by Incremental Moment MatchingNeurIPS 2017
- More Classifiers, Less Forgetting: A Generic Multi-classifier Paradigm for Incremental LearningECCV 2020
- Accelerating Certified Robustness Training via Knowledge TransferNeurIPS 2022
- How Stable Are Transferability Metrics Evaluations?ECCV 2022
- Overcoming Catastrophic Forgetting beyond Continual Learning: Balanced Training for Neural Machine TranslationACL 2022
- Continual evaluation for lifelong learning: Identifying the stability gapICLR 2023
Related open problems
Continual Learning & Catastrophic Forgetting
Characterizing and Reducing Pre-Trained ViT Dependence in Transfer-Based Methods
A wide range of methods built on frozen ImageNet-pretrained ViT features—whether for adapting to new tasks, constraining representations, or detecting anomalies—inherit an unexamined dependency: their guarantees hold only when the backbone's feature space is already good. Today nobody knows whether these methods degrade gracefully with weaker pretraining, generalize to domains far from ImageNet, or transfer to non-transformer architectures, because the dependence is never ablated. The consequence is that reported gains may be properties of the backbone, not of the methods themselves, and the methods are silently inapplicable wherever a strong ImageNet ViT does not exist.
Continual Learning & Catastrophic Forgetting
Sub-Quadratic Covariance Sketching for Exemplar-Free Continual Learning
Exemplar-free continual learning methods increasingly rely on storing per-class feature statistics (prototypes and full covariance matrices) to generate pseudo-replay features or parameterize Mahalanobis classifiers without retaining raw inputs. However, storing a dense $D \times D$ covariance matrix for each of $C$ classes incurs $\mathcal{O}(C \cdot D^2)$ storage, which quickly exceeds device memory when scaling to large label spaces ($C \ge 1000$) or modern high-dimensional embeddings ($D \ge 1024$). Practitioners currently resort to naive diagonal approximations that discard inter-feature correlations and cause severe classification degradation under continual distribution shifts. As a result, covariance-based exemplar-free learning is structurally blocked from scaling to realistic large-vocabulary benchmarks on memory-constrained hardware.
Continual Learning & Catastrophic Forgetting
Task-Agnostic Inference for Modular and Masked Continual Learning Architectures
Modular continual learning methods prevent catastrophic forgetting by allocating task-specific masks, sub-networks, or adaptive parameters during training. However, these methods structurally depend on having ground-truth task identifiers provided at test time to select the corresponding parameters. In practical deployments, incoming inputs arrive without task metadata, rendering task-dependent parameter masking unusable in class-incremental or task-agnostic settings. Without a reliable mechanism to resolve task identity or route inputs dynamically at test time, parameter-isolation architectures remain restricted to artificial task-incremental benchmarks.