Search code, repositories, users, issues, pull requests...

The EfficientNet paper taught us that simply scalling neural networks isn't the most efficient way to generalise to a function and was able to greatly scale down model sizes and outperform state of the art models. Neural golfing challenges are hopefully a fun method of demonstrating the importance of compressed model sizes.

PyTorch's MNIST example produces a model with 1,199,882 parameters that is able to get an accuracy >99%, The current most compressed model in this repo is able to achieve 1% below that accuracy with only 1055 parameters.

Leave a Comment