This is getting more attention than expected, so full acknowledgements: thanks to Christopher Mattern (from @DeepMindAI) who mentioned this to me about two years ago over Friday Drinks as fun fact and to @owencm for a random afternoon conversation turning into a tiny project 🎉💕

Sure but...45% accuracy is not exactly good. You can get close to 88% with a linear classifier. You can get 95% with nearest-neighbor/L2 distance. No deep learning necessary. But if you want more than 99% without losing your computational shirt, go with ConvNets.

Thanks! That's true🤗I would not recommend anyone use this classifier in seriousness😇 I was surprised it is working this well at all and better than nearest-neighbor on pixel sums. At best, it's a simple proof-of-concept for information-theoretic approaches😊

Nice! For completeness, a link to some of the original classification-by-compression work: cs.waikato.ac.nz/~eibe/pubs/Fra…

Thanks! I had been looking around a bit for similar papers but haven't found much. It seems well-known in the statistical compression community. Indeed, I have to thank Christopher Mattern (from @DeepMindAI) for mentioning this over drinks three years ago as a fun fact/idea 😊

I was recently wondering about sth similar: you can probably just count the number of pixels (i.e., just do a sum over the pixel values) to classify MNIST images with ~50% accuracy, which isn't too bad.

Actually, we have tried that 😊You only get 20% accuracy. Zip compression indeed performs significantly better. If you scroll down in the Jupyter Notebook, you can see results for summing on both binarized MNIST and vanilla MNIST. 👉 github.com/BlackHC/mnist_…

Yaa, I mean ummm... It is definitely creative. Btw, isn't even like random coins should get like 50 percent accuracy?

Random baseline accuracy is 10% ☺️

“We are uncertain whether this is an appraisal of zip compression or an indictment of the MNIST dataset.”

After 30 years of optimizing on it, the MNIST test set is no longer a test set; it is rather a validation set.