Abstract: The goal of the paper is to propose "GRAD-MATCH," a data subset selection framework that enhances the efficiency of deep model training by choosing subsets that closely match the gradient of the full training or validation set. By minimizing the gradient matching error, GRAD-MATCH aims to reduce computational and environmental costs associated with large datasets, providing theoretical convergence guarantees and achieving a favorable trade-off between accuracy and efficiency in model training

Abstract: The goal of the paper is to introduce "GLISTER," a framework for efficient and robust learning through data subset selection. GLISTER aims to select a subset of training data that maximizes the model's generalization by optimizing the log-likelihood on a validation set, thus reducing training time and enhancing robustness against noisy and imbalanced data. The approach is formulated as a bi-level optimization problem and is applied across various learning tasks, including active learning

Abstract: The goal of this paper is to propose "CORESET-PFEDBAYES," a Bayesian coreset optimization framework for personalized federated learning. This approach aims to reduce the computational burden of training across distributed clients by leveraging a subset of representative data points, or coresets, from each client. CORESET-PFEDBAYES seeks to improve efficiency and reduce communication while maintaining high accuracy in model updates for federated learning systems​

Abstract: The paper aims to address deep neural networks' tendency to learn spurious feature-label correlations during model distillation. It proposes DEDIER, a novel approach that uses early readouts (predictions from earlier network layers) to identify problematic instances through confident but incorrect predictions. Unlike existing methods that require explicit knowledge of biases, DEDIER automatically detects and corrects biases by modulating the distillation loss based on these early readout signals. The goal is to improve both group fairness measures and overall accuracy of student models. This approach can be easily integrated into any standard model training procedure.

Abstract: The goal of this paper is to introduce Gradient-based Coreset for Robust and Efficient Federated Learning (GCFL), an algorithm that improves the efficiency, robustness, and privacy compliance of Federated Learning (FL) when data is partitioned across clients with limited resources. GCFL achieves this by selecting a coreset of representative data at each client every few communication rounds, reducing computation and energy costs, and enhancing resilience to noisy data. The approach is evaluated on real-world datasets, showing that it significantly reduces computation and communication overhead while maintaining model accuracy in noisy data conditions.

Abstract: The goal of this paper is to introduce AMAL, a method for effectively combining auxiliary losses in supervised learning through a principled approach. AMAL uses a bi-level optimization strategy to learn instance-specific mixing weights for auxiliary losses, leveraging validation data to optimize these weights. This meta-learning framework is designed to enhance model performance in scenarios like knowledge distillation and rule-based denoising. Experiments demonstrate that AMAL achieves performance improvements over existing baselines, providing insights into the mechanisms driving these gains.

Abstract: The goal of this paper is to develop SELCON, an algorithm for efficient data subset selection in L2-regularized regression tasks. SELCON aims to select a subset of training data that allows for faster model training without compromising much on accuracy. The approach leverages a novel formulation that minimizes training loss while maintaining error bounds on a validation set. By reformulating the problem with dual constraints, SELCON benefits from properties like monotonicity and alpha-submodularity, enabling an efficient approximation method. Experiments demonstrate that SELCON achieves a better balance of accuracy and efficiency compared to existing methods.