We at Torus Actions have a deep understanding of deep learning. We are also doing theoretical machine learning research including :

Design of neural networks

Ensembling methods

Translation from sigmoid values to real probabilities

Data normalization, augmentation, segmentation and rebalancing methods.

Loss functions

Fuzzy boundary problems

In particular, we invented new efficient loss functions, topological voting methods, min-max methods for fuzzy boundaries, etc., which helped us achieve more accurate AI models in our projects.