Bakytzhan Kurmanbek

I am interested in the mathematical structure behind training large-scale neural networks, and in particular in where classical optimization theory does and does not carry over to modern deep learning. During my PhD I would like to work on structured matrix parametrizations — Toeplitz, circulant and low-rank families — together with first-order methods such as conditional gradients, to make large-scale training more efficient without losing expressivity. Coming from a mathematics background and several years of industry work on LLM systems, I am especially keen to keep the theoretical side grounded in what actually happens when these models are trained at scale.