ZeroFlow
收藏资源简介:
ZeroFlow是由清华大学和阿里巴巴达摩院联合提出的首个用于评估无梯度优化算法在克服灾难性遗忘方面的基准测试集。该数据集涵盖了多种遗忘场景、模型类型和评估指标,旨在通过前向传播方法探索如何在不依赖梯度信息的情况下缓解灾难性遗忘问题。数据集的内容包括多个任务序列和复杂度的数据集,如CIFAR-100、CUB、ImageNet-A和OmniBenchmark等。通过该基准测试,研究人员揭示了前向传播在管理任务冲突、减少内存需求以及缓解遗忘方面的潜力,并提出了新的优化原则和改进技术。ZeroFlow的应用领域主要集中在持续学习和预训练模型的微调中,旨在解决模型在时间演化数据流中遗忘先前学习任务的问题。
ZeroFlow is the first benchmark dataset jointly proposed by Tsinghua University and Alibaba DAMO Academy for evaluating gradient-free optimization algorithms in overcoming catastrophic forgetting. This dataset covers diverse forgetting scenarios, model types and evaluation metrics, aiming to explore how to mitigate catastrophic forgetting without relying on gradient information via forward propagation methods. The dataset includes multiple task sequences and datasets with varying complexities, such as CIFAR-100, CUB, ImageNet-A and OmniBenchmark. Through this benchmark, researchers have unveiled the potential of forward propagation in managing task conflicts, reducing memory requirements and alleviating forgetting, and put forward novel optimization principles and improved techniques. The application scenarios of ZeroFlow mainly concentrate on continual learning and fine-tuning of pre-trained models, targeting the problem where models forget previously learned tasks in temporally evolving data streams.

- 1ZeroFlow: Overcoming Catastrophic Forgetting is Easier than You Think清华大学, 阿里巴巴达摩院 · 2025年



