IAS Seminar "Towards AI capable of solving long-standing open problems in research"

Speaker:

Julia Sidorova, Computer Science and Engineering, Blekinge Institute of Technology (BTH), Sweden

Abstract:

The talk will reflect my recent research in theory and applications of machine learning with academia and industry (Ericsson, Sony, Boeing, and Telenor). Traditionally the communities of symbolic and numeric machine learning coexisted in parallel, which has changed in the past few years due to breakthroughs in theoretical machine learning.

Numeric data are fed to the classification algorithm in the form of a fixed size vector (in emotion recognition from voice and airline disruption management). Sequential data take a form of a sequence of a variable length (chemicals, genome, and trajectories of mobile phone carriers). My interests lie in making machine learning more generic, intelligent and autonomous. To this end,

1) I have proposed the classification designs that are close to the underlying cognitive and molecular mechanisms and found the general laws of structure that permit to have the same classification pipeline in chemistry and natural language.

2) To make the system capable of autonomous scientific research, I proposed a coherent strategy of automatic feature learning from structured sequences with a subsequent human-readable classification, to convince a practitioner to trust the model.

3) To aid neurological diagnostics, my group have empowered medical AI with the capacity similar to that of human perception of voice emotion.

The new challenge is constructing a "superbrain" for multi-factor medical analysis to shed light on open problems and provide principally new health services. Although mainly thought for medical research, such a superbrain will be able to "think" about other problems, adjustment of lightning in office buildings is included into the sphere of its "duties".

Julia Sidorova was invited by Dr. Jenia Jitsev (JSC).