Dataset Condensation is a newly emerging technique aiming at learning a tiny dataset that captures the rich information encoded in the original dataset. As the size of datasets contemporary machine learning models rely on becomes increasingly large, condensation methods become a prominent direction for accelerating network training and reducing data storage. Despite numerous methods have been proposed in this rapidly growing field, evaluating and comparing different condensation methods is non-trivial and still remains an open issue. The quality of condensed dataset are often shadowed by many critical contributing factors to the end performance, such as data augmentation and model architectures. The lack of a systematic way to evaluate and compare condensation methods not only hinders our understanding of existing techniques, but also discourages practical usage of the synthesized datasets. This work provides the first large-scale standardized benchmark on Dataset Condensation. It consists of a suite of evaluations to comprehensively reflect the generability and effectiveness of condensation methods through the lens of their generated dataset. Leveraging this benchmark, we conduct a large-scale study of current condensation methods, and report many insightful findings that open up new possibilities for future development. The benchmark library, including evaluators, baseline methods, and generated datasets, is open-sourced to facilitate future research and application.

Dataset Condensation is a newly emerging technique aiming at learning a tiny dataset that captures the rich information encoded in the original dataset.

Approximate machine unlearning (AMU) enables models to `forget' specific training data through specialized fine-tuning on a retained (and forget) subset of training set. However, processing this large retained subset still dominates computational runtime, while reductions of unlearning epochs also remain a challenge. In this paper, we propose two complementary methods to accelerate arbitrary classification-oriented AMU method. First, \textbf{Blend}, a novel distribution-matching dataset condensation (DC), merges visually similar images with shared blend-weights to significantly reduce the retained set size. It operates with minimal pre-processing overhead and is orders of magnitude faster than state-of-the-art DC methods. Second, our loss-centric method, \textbf{Accelerated-AMU (A-AMU)}, augments the AMU objective to quicken convergence. A-AMU achieves this by combining a steepened primary loss to expedite forgetting with a differentiable regularizer that matches the loss distributions of forgotten and in-distribution unseen data. Our extensive experiments demonstrate that this dual approach of data and loss-centric optimization dramatically reduces end-to-end unlearning latency across both single and multi-round scenarios, all while preserving model utility and privacy. To our knowledge, this is the first work to systematically tackle unlearning efficiency by jointly designing a specialized dataset condensation technique with a dedicated accelerated loss function. Code is available at https://github.com/algebraicdianuj/DC_Unlearning.

As graph data grows increasingly complicate, training graph neural networks (GNNs) on large-scale datasets presents significant challenges, including computational resource constraints, data redundancy, and transmission inefficiencies. While existing graph condensation techniques have shown promise in addressing these issues, they are predominantly designed for single-label datasets, where each node is associated with a single class label. However, many real-world applications, such as social network analysis and bioinformatics, involve multi-label graph datasets, where one node can have various related labels. To deal with this problem, we extends traditional graph condensation approaches to accommodate multi-label datasets by introducing modifications to synthetic dataset initialization and condensing optimization. Through experiments on eight real-world multi-label graph datasets, we prove the effectiveness of our method. In experiment, the GCond framework, combined with K-Center initialization and binary cross-entropy loss (BCELoss), achieves best performance in general. This benchmark for multi-label graph condensation not only enhances the scalability and efficiency of GNNs for multi-label graph data, but also offering substantial benefits for diverse real-world applications.

Dataset Condensation is a newly emerging technique aiming at learning a tiny dataset that captures the rich information encoded in the original dataset. As the size of datasets contemporary machine learning models rely on becomes increasingly large, condensation methods become a prominent direction for accelerating network training and reducing data storage. Despite numerous methods have been proposed in this rapidly growing field, evaluating and comparing different condensation methods is non-trivial and still remains an open issue. The quality of condensed dataset are often shadowed by many critical contributing factors to the end performance, such as data augmentation and model architectures. The lack of a systematic way to evaluate and compare condensation methods not only hinders our understanding of existing techniques, but also discourages practical usage of the synthesized datasets.

Graph condensation has emerged as an intriguing technique to provide Graph Neural Networks for large-scale graphs with a more compact yet informative small graph to save the expensive costs of large-scale graph learning. Despite the promising results achieved, previous graph condensation methods often employ an entangled condensation strategy that involves condensing nodes and edges simultaneously, leading to substantial GPU memory demands. This entangled strategy has considerably impeded the scalability of graph condensation, impairing its capability to condense extremely large-scale graphs and produce condensed graphs with high fidelity. Therefore, this paper presents Disentangled Condensation for large-scale graphs, abbreviated as DisCo, to provide scalable graph condensation for graphs of varying sizes. At the heart of DisCo are two complementary components, namely node and edge condensation modules, that realize the condensation of nodes and edges in a disentangled manner. In the node condensation module, we focus on synthesizing condensed nodes that exhibit a similar node feature distribution to original nodes using a pre-trained node classification model while incorporating class centroid alignment and anchor attachment regularizers. After node condensation, in the edge condensation module, we preserve the topology structure by transferring the link prediction model of the original graph to the condensed nodes, generating the corresponding condensed edges. Based on the disentangled strategy, the proposed DisCo can successfully scale up to the ogbn-papers100M graph with over 100 million nodes and 1 billion edges with flexible reduction rates. Extensive experiments on five common datasets further demonstrate that the proposed DisCo yields results superior to state-of-the-art counterparts by a significant margin. The source code is available at https://github.com/BangHonor/DisCo.

Dataset Condensation is a newly emerging technique aiming at learning a tiny dataset that captures the rich information encoded in the original dataset. As the size of datasets contemporary machine learning models rely on becomes increasingly large, condensation methods become a prominent direction for accelerating network training and reducing data storage. Despite numerous methods have been proposed in this rapidly growing field, evaluating and comparing different condensation methods is non-trivial and still remains an open issue. The quality of condensed dataset are often shadowed by many critical contributing factors to the end performance, such as data augmentation and model architectures. The lack of a systematic way to evaluate and compare condensation methods not only hinders our understanding of existing techniques, but also discourages practical usage of the synthesized datasets. This work provides the first large-scale standardized benchmark on Dataset Condensation. It consists of a suite of evaluations to comprehensively reflect the generability and effectiveness of condensation methods through the lens of their generated dataset. Leveraging this benchmark, we conduct a large-scale study of current condensation methods, and report many insightful findings that open up new possibilities for future development. The benchmark library, including evaluators, baseline methods, and generated datasets, is open-sourced to facilitate future research and application.