In this web content, the author explores the use of compression techniques for image classification, specifically with the MNIST dataset. They employ GZIP and NCD (Normalized Compression Distance) as tools to measure the complexity and similarity between data points, respectively. The classification is done using k-NN (k-Nearest Neighbors) with the majority class among the k closest neighbors being predicted as the label for a test sample. The author acknowledges that the approach is computationally expensive and only uses a subset of the test images. The provided code demonstrates the implementation of the algorithm. It is worth noting that another article by Andreas Kirsch took a similar approach with around 35% accuracy.
https://jakobs.dev/solving-mnist-with-gzip/