During pretraining, language models exhibit natural ungrokking where learned rules are forgotten based on their support frequency in the corpus, with asymmetric editability of rule survival.
hub
An empirical study of example forgetting during deep neural network learning.arXiv preprint arXiv:1812.05159
26 Pith papers cite this work, alongside 114 external citations. Polarity classification is still indexing.
26
Pith papers citing it
114
external citations · Pith
abstract
Inspired by the phenomenon of catastrophic forgetting, we investigate the learning dynamics of neural networks as they train on single classification tasks. Our goal is to understand whether a related phenomenon occurs when data does not undergo a clear distributional shift. We define a `forgetting event' to have occurred when an individual training example transitions from being classified correctly to incorrectly over the course of learning. Across several benchmark data sets, we find that: (i) certain examples are forgotten with high frequency, and some not at all; (ii) a data set's (un)forgettable examples generalize across neural architectures; and (iii) based on forgetting dynamics, a significant fraction of examples can be omitted from the training data set while still maintaining state-of-the-art generalization performance.
hub tools
citation-role summary
background 1
citation-polarity summary
roles
background 1polarities
background 1representative citing papers
MergeProbe forecasts LoRA adapter mergeability from first-few-percent training signals and outperforms interference-aware baselines on retention while adding low overhead on a five-domain benchmark.
Shared chat-template tokens piggyback narrow finetuning behaviors onto out-of-domain queries; regularizing their KV states (TReFT) reduces emergent misalignment and other off-topic generalization.
Three last-layer coefficients derived from gradient interference sources predict the forgetting rank order of past classes in rehearsal-based CIL, with self-induced interference as the strongest predictor.
PODS is a plug-and-play oscillatory data-volume scheduler that alternates low-ratio regularization phases with high-ratio recovery phases to improve data selection efficiency across training tasks.
Temporal correlations from lazy random walks enable efficient SGD learning of k-juntas via temporal-difference loss on ReLU networks, achieving linear sample complexity in d.
Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.
Targeted perturbation of reward-anticipatory units in VLMs induces anhedonia-like effort avoidance and clinical-scale score drops without impairing baseline task competence.
RAHA applies rank-aware hyperbolic alignment to vision-language dataset distillation by enforcing geodesic alignment in the shared low-rank range and regularizing the residual subspace for improved transfer.
TDS uses per-tree prediction trajectories to derive instance difficulty scores that rank errors better than prior hardness measures and improve active learning, selective prediction, and Mondrian conformal prediction on tabular data.
MDM distills vision-language datasets via joint embedding clustering, weight-space model interpolation, and geometry-aware distribution matching on the unit hypersphere.
LiBaGS scores and selects synthetic data near decision boundaries using proximity, uncertainty, density, and validity, with boundary-gap allocation and marginal stopping to improve training accuracy.
Target-aligned data selection via normalized endpoint loss drop on a validation-induced reference path achieves competitive performance with reduced computational overhead.
Exploiting data symmetries boosts k-NN to select near-optimal low-noise subsets from noisy datasets, approaching Bayes-optimal performance in high dimensions, with learned representations aiding partial symmetry knowledge.
Fine-tuning causes architecture-specific forgetting of individual samples that is stochastic across seeds, with ViTs showing more predictable decay than CNNs and class-level patterns that are semantically consistent.
AlignPrune uses a Dynamic Alignment Score from loss trajectories to identify noisy samples more accurately than per-sample loss, improving pruning accuracy by up to 6.3% on noisy benchmarks.
Data Agent learns a co-evolving sample selection policy end-to-end that accelerates training by over 50% on ImageNet-1k and MMLU with no performance loss.
Electronic structure datasets across materials show high redundancy from low intrinsic dimensionality, allowing pruning to 1/100th size with preserved chemical accuracy.
EPS uses DCT features to cluster patches by spatial-temporal complexity and adaptively samples from the highest cluster, cutting training patches by 75-91.69% and speeding sampling up to 82.1x versus EMT while claiming preserved quality.
MetaCLIP curates balanced 400M-pair subsets from CommonCrawl that outperform CLIP data, reaching 70.8% zero-shot ImageNet accuracy on ViT-B versus CLIP's 68.3%.
Large-scale standardized benchmarks show state-of-the-art dataset distillation methods do not outperform coreset selection on ImageNet-scale data and have substantially higher construction costs.
OrderDP is a plug-and-play data pruning method that selects a random subset then top-q samples to guarantee unbiased surrogate-loss training with convergence analysis and over 40% training cost reduction on CIFAR and ImageNet.
BRAL-T uses TrustSet-guided reinforcement learning for batch active learning and reports state-of-the-art results on 10 image classification benchmarks plus 2 fine-tuning tasks.
LiLAW learns to weight samples as easy, moderate or hard using three global scalars updated by one gradient step on a validation batch to improve noisy training performance.
citing papers explorer
-
Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining
During pretraining, language models exhibit natural ungrokking where learned rules are forgotten based on their support frequency in the corpus, with asymmetric editability of rule survival.
-
Predicting Mergeability of Parameter-Efficient Fine-Tuning Updates
MergeProbe forecasts LoRA adapter mergeability from first-few-percent training signals and outperforms interference-aware baselines on retention while adding low overhead on a five-domain benchmark.
-
The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment
Shared chat-template tokens piggyback narrow finetuning behaviors onto out-of-domain queries; regularizing their KV states (TReFT) reduces emergent misalignment and other off-topic generalization.
-
Understanding Imbalanced Forgetting in Rehearsal-Based Class-Incremental Learning
Three last-layer coefficients derived from gradient interference sources predict the forgetting rank order of past classes in rehearsal-based CIL, with self-induced interference as the strongest predictor.
-
Beyond What to Select: A Plug-and-play Oscillatory Data-Volume Scheduling for Efficient Model Training
PODS is a plug-and-play oscillatory data-volume scheduler that alternates low-ratio regularization phases with high-ratio recovery phases to improve data selection efficiency across training tasks.
-
The Benefits of Temporal Correlations: SGD Learns k-Juntas from Random Walks Efficiently
Temporal correlations from lazy random walks enable efficient SGD learning of k-juntas via temporal-difference loss on ReLU networks, achieving linear sample complexity in d.
-
Eliciting Latent Predictions from Transformers with the Tuned Lens
Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.
-
Reward Valuation in Vision Language Models: Causal Mechanisms Underlying Anhedonia
Targeted perturbation of reward-anticipatory units in VLMs induces anhedonia-like effort avoidance and clinical-scale score drops without impairing baseline task competence.
-
Rank-Aware Hyperbolic Alignment for Vision-Language Dataset Distillation
RAHA applies rank-aware hyperbolic alignment to vision-language dataset distillation by enforcing geodesic alignment in the shared low-rank range and regularizing the residual subspace for improved transfer.
-
Trajectory-Based Difficulty Scoring for Reliable Learning on Tabular Data
TDS uses per-tree prediction trajectories to derive instance difficulty scores that rank errors better than prior hardness measures and improve active learning, selective prediction, and Mondrian conformal prediction on tabular data.
-
Multimodal Distribution Matching for Vision-Language Dataset Distillation
MDM distills vision-language datasets via joint embedding clustering, weight-space model interpolation, and geometry-aware distribution matching on the unit hypersphere.
-
LiBaGS: Lightweight Boundary Gap Synthesis for Targeted Synthetic Data Selection
LiBaGS scores and selects synthetic data near decision boundaries using proximity, uncertainty, density, and validity, with boundary-gap allocation and marginal stopping to improve training accuracy.
-
Let the Target Select for Itself: Data Selection via Target-Aligned Paths
Target-aligned data selection via normalized endpoint loss drop on a validation-induced reference path achieves competitive performance with reduced computational overhead.
-
Leveraging Data Symmetries to Select an Optimal Subset of Training Data under Label Noise
Exploiting data symmetries boosts k-NN to select near-optimal low-noise subsets from noisy datasets, approaching Bayes-optimal performance in high dimensions, with learned representations aiding partial symmetry knowledge.
-
Not All Forgetting Is Equal: Architecture-Dependent Retention Dynamics in Fine-Tuned Image Classifiers
Fine-tuning causes architecture-specific forgetting of individual samples that is stochastic across seeds, with ViTs showing more predictable decay than CNNs and class-level patterns that are semantically consistent.
-
Beyond Loss Values: Robust Dynamic Pruning via Loss Trajectory Alignment
AlignPrune uses a Dynamic Alignment Score from loss trajectories to identify noisy samples more accurately than per-sample loss, improving pruning accuracy by up to 6.3% on noisy benchmarks.
-
Data Agent: Learning to Select Data via End-to-End Dynamic Optimization
Data Agent learns a co-evolving sample selection policy end-to-end that accelerates training by over 50% on ImageNet-1k and MMLU with no performance loss.
-
Surprisingly High Redundancy in Electronic Structure Data Across Materials Explained by Low Intrinsic Dimensionality
Electronic structure datasets across materials show high redundancy from low intrinsic dimensionality, allowing pruning to 1/100th size with preserved chemical accuracy.
-
EPS: Efficient Patch Sampling for Video Overfitting in Deep Super-Resolution Model Training
EPS uses DCT features to cluster patches by spatial-temporal complexity and adaptively samples from the highest cluster, cutting training patches by 75-91.69% and speeding sampling up to 82.1x versus EMT while claiming preserved quality.
-
Demystifying CLIP Data
MetaCLIP curates balanced 400M-pair subsets from CommonCrawl that outperform CLIP data, reaching 70.8% zero-shot ImageNet accuracy on ViT-B versus CLIP's 68.3%.
-
Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets?
Large-scale standardized benchmarks show state-of-the-art dataset distillation methods do not outperform coreset selection on ImageNet-scale data and have substantially higher construction costs.
-
OrderDP: A Theoretically Guaranteed Lossless Dynamic Data Pruning Framework
OrderDP is a plug-and-play data pruning method that selects a random subset then top-q samples to guarantee unbiased surrogate-loss training with convergence analysis and over 40% training cost reduction on CIFAR and ImageNet.
-
Labeled TrustSet Guided: Batch Active Learning with Reinforcement Learning
BRAL-T uses TrustSet-guided reinforcement learning for batch active learning and reports state-of-the-art results on 10 image classification benchmarks plus 2 fine-tuning tasks.
-
LiLAW: Lightweight Learnable Adaptive Weighting to Learn Sample Difficulty & Improve Noisy Training
LiLAW learns to weight samples as easy, moderate or hard using three global scalars updated by one gradient step on a validation batch to improve noisy training performance.
-
Few-Medoids: An Embarrassingly Simple Coreset Selection Method for Few-Shot Knowledge Distillation
Selecting per-class medoids (samples with lowest average L2 distance to all same-class samples in teacher feature space) consistently outperforms random, herding, and k-center Greedy baselines for few-shot knowledge distillation on from-scratch students.
-
Using Explainability as a Training-Time Reliability Signal for Efficient ECG Classification
ERTS applies explanation quality from Grad-CAM to select reliable samples in progressive data dropout for ECG classification, reporting higher macro-F1 and lower training cost across three datasets and multiple architectures.