|
|
With A. Steinberg (UofT), I was first exposed to the physical limits of information processing in quantum information science. Inspired by the science of successful forays of physicists into biology, I realized neuroscience offers a fascinating area to articulate this idea. With F. Wolf at the Max Planck Institute of Dynamics and Self-organization, I focused my attention on the dynamics of neural circuits. I then rounded out my biology and data analysis while developing inference frameworks for probabilistic models of adaptive immune repertoires with A. Walczak and T. Mora at ENS' Physics department in Paris.
Emerging from this training with the sense that the dynamics of inference serves as a powerful lens to understand learning systems, I returned to Canada in 2018 on an IVADO post-doctoral fellowship to work on dynamical systems approaches to deep learning in recurrent neural networks supervised by Y. Bengio (UdeM) and G. Lajoie (UdeM). It became clear to me that forming good representations is intricately linked to what decisions they serve. I started a collaboration with P. Cisek (UdeM) in which I used bounded rational RL agent models to provide the first reward-rate optimal solution of the decades-old 'tokens task' studied in his lab and to show the particular way in which the macaques in the dataset I analyzed behave suboptimally on the task. This normative approach ignores their social milieu: an attentive, yet controlling human providing juice according to an opaque rule varying over longer timescales in a non-transparent way. Stepping towards realism, I articulated a heuristic strategy more robust to an adversary and that gives a quantitative match to behaviour and a qualitative match to neural activity used to validate the fitted models. This outcome reinforced to me the importance of acknowledging the social factors at play in any given decision-making context.
Wanting a deeper understanding of reinforcement learning theory, I joined the Excellence Research Cluster on Autonomous AI led by Irina Rish. Here, I focused on understanding the sources of limitations of prominent RL algorithms in the non-stationary, multi-task setting of continual learning, especially in multi-agent systems, and have started worked on possible means to overcome them.
Transitioning from Neuroscience to Climate Research
For other reasons, during this time, I starting learning about climate change (documented here). I realized how serious the problem is and how, with my skillset of data-driven modelling of multi-agent learning systems, I might contribute. I identified text as a key data source and topic modelling was a natural approach since the models are actually quite similar to (because they were in fact first published in!) genetic sequence models and generative modelling work I had done in immune repertoires. This got me interested in the text-as-data paradigm of computational social science. Combined with multi-agent systems, there are more than enough hard, important problems to pursue for the rest of my career through my desire to work on social dilemmas like climate change where social factors can influence collective and individual beliefs. I hope that this work both articulates the challenge of and opportunity to enhance communication on these social dilemmas.
