Mathematics of Modern Machine Learning (M3L)

Deep learning has demonstrated tremendous success in the past decade, sparking a revolution in artificial intelligence.

However, the modern practice of deep learning remains largely an art form, requiring a delicate combination of guesswork and careful hyperparameter tuning. This can be attributed to the fact that classical machine learning theory fails to explain many deep learning phenomena, which inhibits its ability to provide effective guidance in practice.

As we enter the large model era of deep learning, this issue becomes even more critical since trial and error with billion- or trillion-size models can result in enormous costs of time and computation. There is a greater need than ever before for theory that can guide practice and provide principled ways to train large models...

This M3L workshop explores theories for understanding and advancing modern ML practices.