CIFAR-100

CIFAR-100是由Alex Krizhevsky, Vinod Nair和Geoffrey Hinton整理发布的经典计算机视觉数据集，现已成为评估连续学习（Continual Learning）和少样本连续学习（Few-Shot Continual Learning）算法性能的核心基准之一。在深度学习研究中，它通常被视为比CIFAR-10更具挑战性的任务。在连续学习背景下，由于其包含100个类别，能够被划分为更长的任务序列（如10个或20个连续任务），从而有效地模拟数据流中新类别不断出现的过程，用于评估模型克服“灾难性遗忘”的能力。在少样本连续学习（FSCIL）设置中，它常被用于测试模型在仅有少量样本（如每类5张图片）的情况下，能否在保留旧知识的同时快速适应新类别的能力。该数据集是从大规模互联网图片库中筛选、清洗并调整尺寸构成。 主要内容为32×32像素的彩色RGB图像。 在连续学习实验设置中，研究者通常不使用原始的静态混合数据，而是将其人工划分为多个互斥的任务子集（例如：将100类平均分为10个任务，每个任务包含10个新类别），以此构建动态的增量学习环境。CIFAR-100共包含60,000张图像，其中50000张为训练集，10000张为测试集。 它拥有100个细分类别（如海狸、海豚、水獭、海豹、鲸鱼等），这100个类别又被归纳为20个超类（Superclasses，如水生哺乳动物）。 每个类别包含600张图像（500张训练，100张测试）。 在少样本连续学习的标准协议中，通常将前60类作为基础任务（Base Task），剩余40类作为增量任务（New Tasks），以N-way K-shot（如5-way 5-shot）的形式逐步提供给模型。

CIFAR-100 is a classic computer vision dataset curated and released by Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton, and has now become one of the core benchmarks for evaluating the performance of Continual Learning (CL) and Few-Shot Continual Learning (FSCIL) algorithms. In deep learning research, it is generally regarded as a more challenging task than CIFAR-10. In the context of continual learning, with its 100 categories, it can be divided into longer task sequences (e.g., 10 or 20 consecutive tasks), effectively simulating the process of new categories continuously emerging in data streams, and is used to evaluate the ability of models to overcome catastrophic forgetting. In the Few-Shot Continual Learning (FSCIL) setting, it is often used to test whether a model can rapidly adapt to new categories while retaining old knowledge when only a small number of samples are available (e.g., 5 images per category). This dataset is constructed by screening, cleaning, and resizing images from large-scale internet image repositories, and its main content consists of 32×32 pixel color RGB images. In continual learning experimental settings, researchers usually do not use the original static mixed data, but manually divide it into multiple mutually exclusive task subsets (e.g., evenly split the 100 categories into 10 tasks, each containing 10 new categories) to construct a dynamic incremental learning environment. CIFAR-100 contains a total of 60,000 images, of which 50,000 are training set samples and 10,000 are test set samples. It has 100 fine-grained categories (such as beaver, dolphin, otter, seal, whale, etc.), and these 100 categories are grouped into 20 superclasses (e.g., aquatic mammals). Each category contains 600 images (500 for training and 100 for testing). In the standard protocol for Few-Shot Continual Learning, the first 60 categories are usually used as the Base Task, while the remaining 40 categories are used as New Tasks, which are gradually provided to the model in the form of N-way K-shot (e.g., 5-way 5-shot).

搜集汇总

背景与挑战

背景概述

CIFAR-100是一个包含60,000张32×32像素彩色图像的经典计算机视觉数据集，涵盖100个细分类别和20个超类。它广泛用于评估连续学习和少样本连续学习算法的性能，特别是在模拟数据流中新类别出现和克服'灾难性遗忘'方面。