Ferenc Huszár

Our lab at Cambridge studies foundational aspects of learning and inference, with particular focus on modern deep learning methods.

PhD students

• Annabelle Carrell: calibration in deep neural networks
• Nitarshan Rajkumar: AI safety and governance; AI Safety Institute.
• Siyuan Guo: co-sup with Bernhard Schölkopf, causality, probability, exchangeability
• Kamilė Stankevičiūtė: self-supervised learning, pharmacology, tabular deep learning
• Szilvia Ujváry: LLM theory, in-context learning, implicit Bayesian inference
• Anna Mészáros: LLM theory, rule extrapolation, formal languages
• Patrik Reizinger: co-sup with Wieland Brendel, causal representation learning, identifiability
• Iulia Duta co-sup with Pietro Lió, graph, hypergraph and geometric deep learning
• Royson Lee: co-sup with Nic Lane, federated learning, imbalanced data, meta-learning

Research Assistants

• Euan Ong: differentiable algorithms, mechanistic interpretability, understanding deep learning
• Viktor Mirjanic: co-sup with Challenger Mishra, ML for string theory, conjecture generation
• Binjie (Anya) Chen: co-sup with Challenger Mishra, ML for string theory, surrogate models

Alumni and friends

• Undergraduate: Viktór Tóth, Viktória Csizmadia, Euan Ong, Bence Hervay, Anna Kerekes, Anna Mészáros, Szilvia Ujváry, Zsigmond Telek
• Masters: Ross Viljoen, Ren Cheong, Mina Remeli, Andrei Alexandru, Charlie Tan, Rudolf Laine

For prospective students

I invite PhD applications relating to theory of deep learning: studying learning dynamics in neural networks, characterising and understanding rational behaviour and extrapolation phenomena in autoregressive language models, studying neural network behaviour in algorithmic/mathematical datasets.

Topics I don't currently seek new students in: computer vision, superresolution, image compression, recommender systems, self-supervised representation learning, causal inference.

About me

I'm Ferenc Huszár, Associate Professor of Machine Learning at the University of Cambridge. I joined the Department of Computer Science and Technology, a nice and cozy department, in 2020. I'm interested in principled deep learning techniques: optimization, generalization, representation, transfer, meta-learning, and so on. I focus more on understanding than on developing new techniques.

I did my PhD in Bayesian machine learning with Carl Rasmussen, Máté Lengyel and Zoubin Ghahramani over at the Engineering Department in Cambridge. I worked on topics of approximate inference [1, 2], active learning [3, 4], and applications of these to sciences [5, 6].

Following my PhD I worked in various jobs in the London tech/startup sector. My highlight as a researcher is joining Magic Pony Technology, a startup where we developed deep learning-based image superresolution [7, 8] and compression [9] techniques. After Twitter's acquisition Magic Pony, I have worked on a range of ML topics, like recommender systems [10, 11] and fair machine learning.

inFERENCe

I started this blog some time ago but now it kind of has its own life. inFERENCe got started when, in 2015, I returned to machine learning research after a 3-year stint as a data scientist. I basically slept through the deep learning revolution. In those three years, many things happened, so I had to play catch up.

Initially, these blog posts helped me understand the body of literature I have missed: generative adversarial networks, variational autoencoders, representation learning, etc. Nowadays, I continue reading and writing about current papers, trying to reinterpret them and find connections to things I know, usually ending up with some kind of KL divergence.

If you're new here and want to get a taste, you may want to start with these crowd favourites:

• introduction to causal inference and do-calculus and follow-up posts.
• dilated convolutions and Kronecker products
• Gaussian Distributions as Soap Bubbles
• Everything that works works because it's Bayesian
• GANs are broken in more than one way

Contact

You can send me an email to fh277@cam.ac.uk, follow me on twitter @fhuszar, or look at my Google scholar profile.