Alexandra Maria Proca

Imperial College London, Department of Computing

prof_pic2.jpg

I’m a PhD student and a President’s scholar in the Department of Computing at Imperial College London, supervised by Pedro Mediano and Murray Shanahan. Broadly, my research interests span the fields of computational/theoretical neuroscience and machine learning/mechanistic interpretability. I’m interested in using the tools of machine learning for developing general theories of learning and cognition to better understand both biological and artificial minds. I’m particularly interested in how information is learned, represented, and processed in neural populations (mixed selectivity, superposition, neural dynamics, learning dynamics) for flexible behavior. My work is grounded in studying simple, interpretable models (such as linear networks and low-rank RNNs) using mathematics, statistical physics, and dynamical systems approaches.

Before attending Imperial, I received my bachelors degree in computer science and neuroscience (with a music minor) from the University of North Carolina at Chapel Hill and then completed a masters degree in machine learning at University College London. During my degrees, I worked as a research assistant in several labs on various topics in the fields of machine learning and neuroscience. After finishing my masters, I worked as a research assistant in the Department of Computer Science at ETH Zürich with João Sacramento, studying the use of hypernetworks for meta-learning. For more information, you can view my CV.

I really enjoy discussing and engaging with science and philosophy with other people. I currently help organize Qualiaheads, a club of graduate students studying the state of research in consciousness science.

Outside of research, I love anything outdoors (marathon running, hiking, skiing, etc.). I also enjoy playing music and writing. I’ve been playing piano for 18 years and while I lived in Zürich, I was a singer in a local band. I occassionally write poetry and (less frequently) share it.

news

Oct 2024 Our commentary paper on supporting NeuroAI trainees is now out in Nature Communications.
Oct 2024 Our paper on rich and lazy learning was accepted at the NeurIPs UniReps Workshop.
Sep 2024 Our paper on task abstractions in gated linear networks was accepted as a NeurIPs Spotlight!
Aug 2024 I’ll be presenting a poster on learning dynamics in linear RNNs at CCN in Boston.
Jun 2024 I’m giving a talk on context abstractions in linear networks at the ICTP Workshop in Trieste, Italy
May 2024 Our paper on PID in multitask ANNs was accepted to PLoS Computational Biology.
Jan 2024 I’ll be presenting our work on inferring context in gated linear networks at COSYNE.
Jan 2024 Our paper on discovering modular solutions that generalize compositionally was accepted to ICLR.

selected publications

  1. Flexible task abstractions emerge in linear networks with fast and bounded units
    NeurIPS Spotlight, 2024
  2. From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks
    NeurIPS UniReps Workshop, 2024
  3. Discovering modular solutions that generalize compositionally
    ICLR, 2024
  4. Synergistic information supports modality integration and flexible learning in neural networks solving multiple tasks
    PLoS Computational Biology, 2024