The paper analyzes how forgetting in LLMs is linked to simple task-upstream example associations.
problem Forgetting of upstream knowledge in fine-tuned LLMs.
method Empirical analysis of forgotten examples in N upstream examples after M new tasks, using low-rank matrix approximation. result Forgetting can be predicted efficiently using matrix completion over empirical associations.
Paper benchmarks CF mitigation in federated time series forecasting.
problem Catastrophic forgetting in federated learning for time series forecasting.
method Comprehensive evaluation of CF mitigation strategies in federated time series forecasting.
result Introduction of a new benchmark for CF in time series federated learning.
Our method upweights easy samples to mitigate forgetting in fine-tuning.
problem Catastrophic forgetting in fine-tuning pre-trained models.
method Sample weighting based on pre-trained model's losses.
result Our method reduces forgetting by up to 0.8% on MetaMathQA while preserving more accuracy on pre-training datasets.
This paper investigates how forgetting affects neural network representations and stabilizes deeper layers.
problem Catastrophic forgetting in machine learning models trained on sequential tasks.
method Representational analysis techniques and empirical studies on CIFAR-10 and CIFAR-100 datasets.
result Deeper layers are disproportionately the source of forgetting, and methods to mitigate forgetting stabilize these layers.
Paper proposes a method to maintain ASR performance on new tasks without forgetting old ones.
problem Mitigating forgetting in ASR models when learning new tasks.
method A novel explainability-based knowledge distillation combined with response-based knowledge distillation.
result Our method outperforms existing ones in mitigating forgetting on multi-stage sequential training tasks.
Overparameterization helps prevent forgetting in sequential learning tasks.
problem Catastrophic forgetting in continual learning systems.
method Analytical study of gradient descent with linear regression model.
result Overparameterization can mitigate forgetting in a two-task setting.
Proposes a method to adapt DNNs to drift in data distribution.
problem Adapting to out-of-distribution data and shifting objectives.
method Bayesian Inference, Variational Density Propagation, Evidence Lower Bound (ELBO), Minimum Description Length (MDL) Principle.
result Minimizes catastrophic forgetting by approximating MDL principle.
EVCL combines VCL and EWC to prevent forgetting new tasks.
problem Preventing catastrophic forgetting in continual learning.
method Hybrid model integrating VCL and EWC.
result Consistently outperforms baselines in learning new tasks.
This work tackles catastrophic forgetting in neural networks by mimicking brain's metaplasticity.
problem Catastrophic forgetting in neural networks, where new tasks erase previously learned ones.
method Interpreting binarized neural networks as metaplastic systems, adjusting their training technique.
result Training technique reduces catastrophic forgetting without needing previously presented data.
Paper tackles few-shot class-incremental learning with a neural gas network.
problem Incrementally learn new classes from very few labelled samples without forgetting old classes.
method Proposes TOPIC framework using a neural gas network to preserve class topology and adapt to new samples.
result Significantly outperforms other methods on CIFAR100, miniImageNet, and CUB200 datasets.
Dynamic memory prevents forgetting in continuous learning of medical images.
problem Catastrophic forgetting in machine learning models over time due to domain shifts.
method Dynamic memory to store and replay diverse training data subsets.
result Dynamic memory mitigates forgetting without knowing when shifts occur.
Study shows how to balance memory and learning efficiency in continual learning.
problem Balancing memory and learning efficiency in continual learning.
method Structural regularization with Hessian-based regularization.
result Structural regularization improves statistical performance at the cost of increased memory complexity.
Study explores D-Wave QA for RBM training, finds modest benefits but no significant improvements.
problem Improving RBM training using D-Wave quantum annealing.
method Hybrid sampling approach combining classical and D-Wave QA methods.
result No significant improvements in RBM training using D-Wave QA, but potential for CF mitigation.
Improved GANs mitigate forgetting and collapse issues.
problem Catastrophic forgetting and mode collapse in GANs.
method Contrastive learning and mutual information maximization.
result Significantly stabilizes GAN training and improves performance.
LaLoRA prevents forgetting in LoRA fine-tuning.
problem Catastrophic forgetting in fine-tuned models.
method LaLoRA applies Laplace approximation to LoRA weights for regularization.
result Improved learning-forgetting trade-off with controllable regularization strength.
Unified study of stateful replay for streaming learning, reducing forgetting by 2-3x.
problem Catastrophic forgetting in streaming generative and predictive learning.
method Unified analysis of stateful replay for autoencoding, forecasting, and classification tasks.
result Stateful replay reduces average forgetting by a factor of 2-3 on heterogeneous multi-task streams.
Current training regimes for deep learning usually involve exposure to a single task / dataset at a time. Here we start from the observation that in this context the trained model is not given any knowledge of anything outside its (single-task) training distribution, and has thus no way to learn parameters (i.e., featu…
ER-GNN uses experience replay to prevent GNNs from forgetting previous tasks.
problem Catastrophic forgetting in GNNs when learning multiple tasks sequentially.
method Experience Replay framework to store and replay knowledge from previous tasks.
result ER-GNN effectively mitigates catastrophic forgetting in GNNs.
The paper tackles CF in CL by analyzing NTK overlap matrix and proposing OGD.
problem Catastrophic Forgetting in continual learning.
method Analysis of NTK overlap matrix, OGD with PCA.
result Proposes OGD to mitigate CF, supported by experiments.
Proposes a method to retrain neural networks incrementally for continuous data flow.
problem Continuous data flow and the challenges of catastrophic forgetting and efficient retraining.
method Incremental retraining using multi-armed bandits to select important samples and weights, and a new regularization term for synapse and neuron importance.
result Mitigates catastrophic forgetting and boosts model performance.
New insights into continual learning with task similarity.
problem Challenges in learning similar tasks without interference.
method Linear teacher-student model with latent structure.
result High input feature similarity with low readout similarity is catastrophic.
Using neural networks in practical settings would benefit from the ability of the networks to learn new tasks throughout their lifetimes without forgetting the previous tasks. This ability is limited in the current deep neural networks by a problem called catastrophic forgetting, where training on new tasks tends to se…
Paper analyzes fine-tuning methods for machine unlearning, proposing a new strategy to improve forgetting accuracy.
problem Improving fine-tuning methods to effectively forget specific subsets of data in machine learning models.
method Theoretical analysis and a novel Retention-Based Masking (RBM) strategy are proposed.
result RBM significantly improves unlearning accuracy while preserving retaining accuracy.
A new memory replay mechanism improves reinforcement learning stability and speed.
problem Forgetting in reinforcement learning with continuous control.
method Augmented Memory Replay (AMR) that optimizes the replay of past experiences.
result AMR enhances stability and convergence speed of learning algorithms.
New method improves ABI for sequential data, reducing forgetting and improving accuracy.
problem Performance degradation of ABI under model misspecification and distribution shifts.
method Decouples simulation-based pre-training from unsupervised SC fine-tuning, using memory buffer and elastic weight consolidation.
result Significant mitigation of forgetting and improved posterior estimates compared to standard simulation-based training.
This research debiases machine unlearning by using counterfactual examples.
problem Machine unlearning processes can be biased, leading to inaccurate results.
method Intervention-based approach using counterfactual examples to mitigate biases.
result The method outperforms existing baselines on evaluation metrics.
Study examines how image artifacts impact polyp detection and proposes methods to mitigate their effects.
problem Impact of image artifacts on automated polyp detection accuracy.
method Systematic analysis of six artifact classes, investigation of learning without forgetting framework.
result Artifacts can either benefit or harm polyp detection; learning without forgetting can mitigate some harmful effects.
Modern Hopfield networks help prevent forgetting in generative models after task changes.
problem How to prevent forgetting in generative models after task changes.
method Introduce intrinsic forgetting as an increase in Hopfield energy after task change, analyze memory replay effectiveness, and validate predictions in experiments.
result High-energy, outlier-like samples are more forgettable than cluster-like samples, and energy-based selection of replay samples mitigates forgetting.
CILF learns adaptive embeddings for class-incremental learning with novel class detection and model update.
problem Handling unknown classes and model update in streaming data with new classes.
method CILF uses decoupled prototype based loss for intra-class and inter-class structure improvement, and a learnable curriculum clustering operator for adaptive embedding.
result CILF effectively detects multiple novel classes and mitigates embedding confusion, while updating the model without catastrophic forgetting.
FedGTEA learns new tasks in federated learning with task embeddings and alignment.
problem Federated class-incremental learning with task-specific knowledge and model uncertainty.
method Cardinality-Agnostic Task Encoder (CATE) for Gaussian task embeddings, 2-Wasserstein distance for inter-task alignment.
result FedGTEA achieves superior classification performance and mitigates forgetting.
Centroids Matching tackles catastrophic forgetting by matching feature vectors to class centroids.
problem Catastrophic forgetting in neural networks when learning new tasks.
method Centroids Matching operates in the embedding space of neural network features, matching these vectors to class centroids.
result Centroids Matching achieves high accuracy on all tasks without using external memory, even in realistic scenarios.
The condition number predicts efficient information encoding in neural units, aiding model fine-tuning.
problem Efficient information encoding in neural units for various tasks and input modalities.
method Linking the condition number to the log-volume scaling factor and entropy of the output distribution.
result High condition number indicates efficient encoding, reducing overall information transfer.
Efficiently removes specific data subsets without retraining for GDPR compliance.
problem Efficiently removing specific data subsets to comply with GDPR regulations.
method Statistical framework for machine unlearning with minimax optimality for squared loss.
result Developed Unlearning Least Squares (ULS) achieving minimax optimality for estimating model parameters.
Incremental class learning, a scenario in continual learning context where classes and their training data are sequentially and disjointedly observed, challenges a problem widely known as catastrophic forgetting. In this work, we propose a novel incremental class learning method that can significantly reduce memory ove…
ADER addresses continual learning in session-based recommendation by periodically replaying exemplars with adaptive distillation.
problem Catastrophic forgetting in continual learning of session-based recommenders.
method Periodically replaying previous training samples (exemplars) with an adaptive distillation loss.
result ADER consistently outperforms other continual learning techniques and even all historical data at every update cycle.
Framework for safely updating machine learning models.
problem Continuous updates to machine learning models can lead to unintended consequences.
method Formalizes the problem as computing the largest locally invariant domain (LID), uses tractable primal-dual formulation.
result Matches or exceeds heuristic baselines for avoiding forgetting while providing formal safety guarantees.
Experience replay (ER) is a fundamental component of off-policy deep reinforcement learning (RL). ER recalls experiences from past iterations to compute gradient estimates for the current policy, increasing data-efficiency. However, the accuracy of such updates may deteriorate when the policy diverges from past behavio…
Paper tackles class-incremental learning by combining self-supervised learning to mitigate prior information loss.
problem Catastrophic forgetting and prior information loss in class-incremental learning.
method Combining self-supervised learning with class-incremental learning to mitigate prior information loss.
result Proposed method outperforms state-of-the-art methods.
New critics improve VAEs by preventing latent collapse.
problem Posterior collapse in VAEs where latent variables are ignored.
method Inference critics that detect and incentivize meaningful latent representations.
result Optimizing inference critics increases mutual information between latent and observations.
Class incremental learning refers to a special multi-class classification task, in which the number of classes is not fixed but is increasing with the continual arrival of new data. Existing researches mainly focused on solving catastrophic forgetting problem in class incremental learning. To this end, however, these m…
A simple technique improves continual learning by 50% on image datasets.
problem Challenges in training neural networks on a stream of shifting data.
method Experience Replay (ER) with five tricks to mitigate its shortcomings.
result ER, enhanced with tricks, achieves significant accuracy gains.
Research shows continual learning challenges in confounded datasets.
problem Challenges in mitigating confounders in continual learning settings.
method Formal description of continual confounders, construction of ConCon dataset.
result Standard continual learning methods fail to ignore confounders.
A new method for continual learning in GANs learns new modes with limited data.
problem Learning new target modes with limited samples while preserving previously learned ones.
method Mode-affinity score for generative modeling, generator replay, and weighted label generation.
result Gains over state-of-the-art methods, even with fewer training samples.
DCL method improves congruency in machine learning tasks.
problem Varying task-specific semantics in training data.
method Direction Concentration Learning (DCL) method to enhance congruency.
result Improves performance in saliency prediction, continual learning, and classification tasks.
La-MAML improves fast online continual learning with a look-ahead approach.
problem Fast online continual learning with limited model capacity.
method Optimisation-based meta-learning with look-ahead and episodic memory.
result Superior performance on visual classification benchmarks.
Continual learning aims to learn new tasks without forgetting previously learned ones. This is especially challenging when one cannot access data from previous tasks and when the model has a fixed capacity. Current regularization-based continual learning algorithms need an external representation and extra computation …
A new framework ensures model safety by retaining old model capabilities while improving new tasks.
problem Catastrophic forgetting in continual learning systems.
method Retention-centric framework with data-dependent constraints and an efficient constrained optimization algorithm.
result The method ensures that new models retain important capabilities of old models while improving new tasks.
CPR adds entropy maximization to improve continual learning methods.
problem Catastrophic forgetting in continual learning.
method Classifier-Projection Regularization (CPR) adds an entropy maximization term to existing regularization methods.
result CPR improves accuracy and plasticity in continual learning methods.