Botos Csaba

mail github linkedin google-scholar twitter cv

Botos Csaba; DPhil in Machine Learning at Univesity of Oxford

About me

The thread tying my work together is teaching, which I think is the overlooked problem in theory of mind: you cannot teach a mind without first modelling how it thinks. So I study how frontier reasoning models form, test, and revise hypotheses, and how their internal representations line up with human cognition, using controlled behavioural and neural experiments. I then ask how that understanding lets machines teach humans, by tracking what a learner knows and choosing the tasks that move them forward, and lets artificial agents teach and coordinate with one another.

[DEMO] I am passionate about communicating ML research to a wide audience and I have made TensorFlow.js visualizations: Domain Partitioning via Adversarial Training, Real-Time Online Continual Learning, Interactive Contrastive Learning.

I have been awarded the Encode Fellowship funded by the UK Government and managed by Pillar VC to work on Adaptive Curricula for Human Learning with Christopher Summerfield and Rui Ponte Costa. In addition, I hold a Junior Research Fellowship at Wolfson College, University of Oxford.

Previously, I defended my DPhil thesis in Engineering Science at University of Oxford, supervised by Prof. Philip Torr and Dr. Adel Bibi and examined by Prof. Ferenc Huszár and Dr. Oiwi Parker Jones. My thesis focuses on compute efficient distribution drift detection and test time adaptation across vision and language tasks.

During my undergraduate studies, I interned at the Institute of Experimental Medicine under the supervision of Dr. Gábor Nyiri, at Verizon Smart Communities under the supervision of Dr. András Horváth. During my DPhil, I was a research intern at the Intel Embodied AI Lab.

Selected Publications

Teaching

I have been hosting and organizing Machine Learning Seminars and Reading Groups:

I have organized outreach events for high school students and been a mentor for: Furthermore, I have been a teaching assistant for the following courses:

Projects

[DEMO] Reason to Play: Do frontier reasoning models learn novel grid-world games the way humans do, given no rules? We benchmark 8 frontier LRMs against 32 fMRI-scanned humans and deep-RL baselines: the strongest models match human rule-discovery and curriculum progression, and their hidden states predict human brain activity during in-context learning. We read the chain-of-thought to relate stated reasoning to the actions actually taken.
reason-to-play
[DEMO] Generative Adversarial Networks Playground: An interactive visualization toolkit for understanding the training dynamics of GANs.
ganplayground
[DEMO] Label Delay in Continual Learning: How do we utilize data from the future tasks before we have access to their labels?
ldocl-1 ldocl-2
[DEMO] Wolfson Scheduler: An open source project building on a novel vectorized solution for multi-set permutation problems.
[DEMO] Contrastive Learning in your browser!: How can we visualize the metric space learned by contrastive methods such that elementary school students can understand it?
embedding
Diversified Dynamic Routing for Vision Tasks: how can we improve the performance of mixture of experts models by explicitly diversifying the experts?
divdr-project
Multilevel Knowledge Transfer for Cross-Domain Object Detection: How can we maximize the transfer of knowledge between different domains using image-level, feature-level and instance-level knowledge transfer?
mkt
Domain Partitioning Network: How can we partition the data distribution efficiently using multiple discriminator networks?
dopanet
Multi-domain classification of cardiac signals with Deep Neural Networks: How to detect cardiac arrhythmia from portable ECG devices?
ecg
SAM: Open-source face recognition framework: Full stack face recognition framework from data collection to model deployment.
sam