CIFAR-100 is a dataset of 60,000 32x32 color images in 100 classes, with 600 images per class. It is a labeled subset of the 80 Million Tiny Images dataset, published in 2009 by researchers at the Canadian Institute For Advanced Research. The dataset is commonly used to train and evaluate machine learning and computer vision algorithms, particularly Deep learning models. Its low resolution allows researchers to quickly test different approaches before scaling to larger datasets.
The 100 classes are grouped into 20 superclasses. For example, the superclass 'aquatic mammals' includes beavers, dolphins, otters, seals, and whales. Each image has two labels: a fine label (the specific class) and a coarse label (the superclass). This structure enables both fine-grained classification and hierarchical learning tasks. The dataset is split into 50,000 training images and 10,000 test images, with each class having 500 training and 100 test images.
History and Creation
CIFAR-100 was introduced alongside CIFAR-10 in 2009. The images were sourced from the 80 Million Tiny Images dataset, a collection of low-resolution images gathered from the web. Students were paid to label the images, ensuring accurate ground truth. The dataset was created to provide a more challenging alternative to CIFAR-10, which has only 10 classes. The higher number of classes increases the difficulty of classification tasks, making CIFAR-100 a standard benchmark for evaluating Neural network architectures.
Usage in Machine Learning
CIFAR-100 is widely used in Machine learning research for tasks such as image classification, object recognition, and transfer learning. Researchers often use it to compare the performance of different Artificial intelligence models. Convolutional neural networks (CNNs) tend to achieve the best results on CIFAR-100, similar to CIFAR-10. The dataset is also used for benchmarking computational efficiency, as teams compete to train models faster and cheaper on hardware like AWS Trainium or Cerebras systems.
Benchmarks and State-of-the-Art
Many research papers report state-of-the-art results on CIFAR-100, but comparisons are complicated by varying preprocessing techniques such as image flipping, shifting, and augmentation. Some models achieve error rates below 10%, but exact rankings depend on experimental setups. The dataset is also part of DAWNBench, a benchmark suite for training and inference performance. As of recent years, Transformer (architecture)-based models and advanced CNNs have shown strong performance, though no single architecture dominates.
Related Datasets
CIFAR-100 is part of a family of similar datasets. CIFAR-10 has 10 classes with 6,000 images each. CIFAR-10H provides human perceptual uncertainty labels for CIFAR-10 images. ImageNet (ILSVRC) contains over one million higher-resolution images across 1,000 classes. The Street View House Numbers (SVHN) dataset has 32x32 color images of digits. The 80 Million Tiny Images dataset serves as the source for both CIFAR variants. These datasets collectively support research in Computer vision and Deep learning.
See Also
- mnist-database
- list-of-datasets-for-machine-learning-research