Efficient training of Neural Networks

Purpose and goal:

This project aimed to develop computationally and energy efficient algorithms for training AI models in PyTorch. The architecture we worked with can be parallelized, which makes the inference even more efficient. An important results is that we now have a demo of efficient algorithms that demonstrate how much better they are than today´s technology. It will be easier to sell in to other actors.

Expected results and effects:

An important results is that we now have a demo of efficient algorithms that demonstrate how much better they are than today´s technology. It will be easier to sell in to other actors.

Approach and implementation:

The project was carried out at Stanford in collaboration with Prof. Mert Pilanci and his group. The goal was to develop our patented efficient AI training algorithms in PyTorch, which we managed to achieve. The performance of the algorithms was verified on open source data.