CIFAR-10 dataset contains 50,000 training images in 10 categories, with a test set of 10,000 images for classification. The state of the art accuracy is 80% achieved using whitening, k-means, and soft activation functions by Adam Coates et al. Human accuracy on the dataset is around 94%, with some challenging images. The dataset shows a wide variety of objects within the same class, making classification challenging. The author’s method for classification involved recognizing distinct parts or using overall cues in the image. The author predicts possible accuracy improvements up to 90%, despite initial doubts about training data size. The post concludes with encouragement for readers to try the classification task themselves for a better understanding of image classification challenges.
http://karpathy.github.io/2011/04/27/manually-classifying-cifar10/