HyperAI超神经

首页资讯最新论文教程数据集百科 SOTA LLM 模型天梯 GPU 天梯顶会

中文

HyperAI超神经

Continual Learning

持续学习（Continual Learning），又称增量学习（Incremental Learning）或终身学习（Life-long Learning），是指在不遗忘先前任务知识的前提下，顺序学习多个任务的模型训练方法。该方法在新任务训练时无法访问旧任务数据，通过提供任务标识（task-id）进行验证。持续学习旨在提升模型在动态环境中的适应能力，具有重要的应用价值，特别是在数据不断变化的场景中。

20Newsgroup (10 tasks)

5-dataset - 1 epoch

CIFAR-100 AlexNet - 300 Epoch

CIFAR-100 ResNet-18 - 300 Epochs

Cifar100 (10 tasks)

Cifar100 (20 tasks)

Model Zoo-Continual

Cifar100 (20 tasks) - 1 epoch

Coarse-CIFAR100

Model Zoo-Continual

CUB-200-2011 (20 tasks) - 1 epoch

CUBS (Fine-grained 6 Tasks)

CondConvContinual

F-CelebA (10 tasks)

CAT (CNN backbone)

Flowers (Fine-grained 6 Tasks)

CondConvContinual

ImageNet-50 (5 tasks)

ImageNet (Fine-grained 6 Tasks)

CondConvContinual

mini-Imagenet (20 tasks) - 1 epoch

MiniImageNet ResNet-18 - 300 Epochs

Model Zoo-Continual

Sketch (Fine-grained 6 Tasks)

Split CIFAR-10 (5 tasks)

split CIFAR-100

Split MNIST (5 tasks)

Stanford Cars (Fine-grained 6 Tasks)

Tiny-ImageNet (10tasks)

TinyImageNet ResNet-18 - 300 Epochs

visual domain decathlon (10 tasks)

Wikiart (Fine-grained 6 Tasks)