Hi all,

Theo will be speaking at tomorrow's group meeting. It will be held at the normal time and place (SS571 on the Toronto side). See below for title and abstract.

Best,

Riley

-----------------------

Title: Making Better Aluminum Alloys with Variational Autoencoders: Success, failures and (my) best practice

Abstract: In this talk, I will give you an overview of a client project that I worked on at IBM. This project involved raw data coming from a factory manfacturing aluminum alloys.

Although the data was abundant, it was noisy and the structure of the data relatively complex.

To overcome these challenges, I built on the idea of the molecular VAE. Instead of having a single VAE, I used two: one to deal with process-related data and one for the chemical composition data. This lead to a very flexible framework where it is possible to optimize the latent of either both networks or of a single one depending on the need.

I will also briefly speak about the challenges of building and training such a network, what I learned and which (hopefully) best practices I came up with.