I am originally from Croatia, where I’ve got my MSc degree in computer science, from the Faculty of Electrical Engineering and Computing (FER) in Zagreb, where I was a part of TakeLab.
During that time I became interested in understanding how neural networks really work, and how one might implement them based on simple and elegant principles. I took the initiative to reimplement a number of neural networks from scratch: Differentiable Neural Computers, Generative Adversarial Networks and Synthetic Gradients. I also implemented own automatic differentiation framework.
In my penultimate year I took a Haskell course, opening my eyes to the fact that programming need not be different than mathematics. This coincided with the release of the paper Backprop as Functor, which made me aware of the field of category theory.
I was hooked, and wanted to learn more. I attended the 0th conference on Applied Category Theory and met the wonderful community behind it. They were extremely welcoming and supportive in my journey to learn. There I’ve met Neil Ghani who invited me to do a PhD in Glasgow under his guidance, to learn about category theory, and apply the ideas behind it to deep learning.
During my PhD, I have written a number of papers exploring the intersection of category theory, machine learning, game theory, and cybernetics, and had a wonderful time in the Mathematically Structured Programming Group at the University of Strathclyde in Glasgow. I’ve learned a great deal about what mathematics and programming are from people working on cutting edge ideas.
On 21st of September 2023 I defended my PhD thesis “Fundamental Components of Deep Learning: A category-theoretic approach”, and passed my viva with minor corrections.
In the following months, I joined the company Symbolica AI, where the research I spearheaded defining the field of Categorical Deep Learning was accepted into ICML and was a key signal for the company’s $31m funding round. Shortly after, I was fired without warning. All the key researchers involved in this work were also fired. This event was covered in a blog post by a member of the category theory community, but unfortunately due to contractual obligations I myself am not allowed to discuss it publicly.
Since, I’ve been involved with industry and research on the research programme of provably correct code generation:
- I founded the company Coend building neural networks that generate provably correct code
- I’m employed by GLAIVE on the research grant building intermediate language for tactics
- I’m a research advisor to Google DeepMind, where I am working with the team led by Petar Veličković
- I’m building TensorType, a type-safe tensor processsing framework in Idris
My CV is here.
Contact
Click to reveal email. You can also find me on Mathstodon, BlueSky and Github.
You can subscribe to my RSS feed.