Maximilian Puelma Touzel

Research Associate at Mila, Quebec
neuroscience, machine learning, dynamics

PhD, Physics (IMPRS Physics of biological and complex systems), University of Göttingen
MSc, Physics, University of Toronto
BSc, Physics & Mathematics, University of Toronto

cebeuq [tod] alim [ta] mtamleup

I am a research associate at Mila affiliated with the Université of Montréal. My research interests lie at the intersection of artificial intelligence/machine learning and neuroscience and the NeuroAI synergy to be found there.

Aug. 2021 Urgency work out as a preprint! Twitter summary thread here.
Jan. 2021 Urgency work peer-reviewed and accepted at COSYNE.
Feb. 2021 Join us for a exciting day of talks and a panel of top experts on the goals and challenges for robust scientific explanations in neural and artificial inteligence systems.
Dec. 2020 Urgency work peer-reviewed and accepted at the Biological and Artificial Reinforcement Learning workshop at NeurIPS.
Dec. 2020 Montreal Artificial Intelligence and Neuroscience 2020 conference.
Nov. 2020 Presented urgency work at the inaugural NeuroAI conference, NAISYS, at Cold Spring Harbor Labs.
May 2020 Happy to lead the breakout session on higher cognition at UNIQUE's inaugural NeuroAI symposium.
Apr. 2020 Check out our preprint on Lyapunov spectra for RNN training.
Apr. 2020 Our work on inferring population dynamics from intrinsically variable, vastly subsampled, and indirectly accessed genetic sequences is published in PLoS Computational Biology.
Mar. 2020 We are presenting our decision-making work in collaboration with the Cisek Lab at COSYNE.
Dec. 2019 We are organizing a NeuroAI workshop at NeurIPS2019.
Dec. 2019 nnRNN paper presented in main track of NeurIPS.
Nov. 2019 With yet another wonderful edition of the Montreal Artificial Intelligence and Neuroscience (MAIN) conference, Montreal is further establishing itself as the hotbed for incisive NeuroAI research. Honoured to have my new postdoc work recognized with two awards here.
May 2019 We are organizing a Physics and AI Workshop in Montreal.
Jan 2019 family_size+=1!


How do we abstract the world in a way that facilitates choosing actions to achieve a goal? How does the uncertainty in our beliefs about the world transform into uncertainty about how valuable are the actions we can take in it? When do we make the effort to reduce this uncertainty and when do we account for it? These questions illustrate the complex interplay between efficiently representating the world and profitably acting in it.

Dynamics (as iterates in an data-driven learning algorithm or evolving neural activity) is central to this interplay. I am interested in how the dynamics of learning agents (be it humans, animals, or machines) implement decision-making. My current interests lie in the intersection of neural dynamics, reinforcement learning, and the cognitive and systems neuroscience of decision-making. I develop theoretical and computational approaches in collaboration with diverse experimental groups.

curriculum vitae [ (noun) latin. Trajectory of one's life.]

With Aephraim Steinberg (UofT), I was first exposed to the physical limits of information processsing 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 Fred Wolf at the Max Planck Institute of Dynamics and Self-organization, I then focussed 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 Aleksandra Walczak and Thierry Mora at the ENS in Paris.

Emerging with the sense that the dynamics of inference serves as a powerful lens to understand learning systems, in late 2018, I returned to Canada on an IVADO post-doctoral fellowship to work on dynamical systems approaches to deep learning in Recurrent Neural Networks (RNNs) supervised by Yoshua Bengio (UdeM) and Guillaume Lajoie (UdeM). It became clear to me that forming good representations is intricately linked to what decisions they serve. I got involved in decision-making research using Mila's excellent learning environment and started a collaboration with the experimental lab of Paul Cisek (UdeM). We've since developed a heuristic learning algorithm, Performance-Gated Deliberation, that approximates average-reward reinforcement learning and is supported by neural dynamics and behaviour of decision-making in non-human primates.

As a mathematical theory of agent-environment interactions, reinforcement learning is also useful when interpreting human decision-making. Many of the pressing societal problems of our time rest on how we make decisions, increasingly while interacting with technology and algorithms. I am currently engaging cognitive neuroscientists and psychologists on the dynamics of human decision-making and how it impacts the AI alignment problem. How can we quantify how we assess risk and how this capacity develops? I am currently collaborating with the developmental and clinical psychology group of Patricia Conrod (UdeM/St. Justine hospital), leading a collaboration on applying inverse reinforcement learning methods to adolescent behavioural data from risk-based decision-making tasks.

I am currently funded under the Canadian Excellence Research Chair in AI held by Irina Rish.

Publications & preprints

Deliberation gated by opportunity cost adapts to context with urgency
Puelma Touzel M, Cisek P, Lajoie G.
preprint | supp. video ]

On Lyapunov Exponents for RNNs: Understanding information propogation using dynamical systems tools
Vogt R*, Puelma Touzel M*, Shlizerman E, Lajoie G.
preprint ]

Inferring the immune response from repertoire sequencing
Puelma Touzel M,Walczak A, Mora T.
PLoS Comp (2020).
journal ]

Non-normal Recurrent Neural Network (nnRNN): learning long time dependencies while improving expressivity with transient dynamics
Goyette K, Kerg GC, Puelma Touzel M, Gidel G, Vorontsov E, Bengio Y, Lajoie G.
NeurIPS 2019.
proceedings ]

Statistical mechanics of spike events underlying phase space partitioning and sequence codes in large-scale models of neural circuits
Puelma Touzel M, & Wolf F.
PRE (2019).
journal | preprint | supp. video ]

Precise tracking of vaccine-responding T-cell clones reveals convergent and personalized response in identical twins
Pogorelyy M, Minervina A, Puelma Touzel M, Sycheva A, Komech E, Kovalenko E, Karganova G, Egorov E, Komkov A, Chudakov D, Mamedov I, Mora T, Walczak A, Lebedev Y.
PNAS (2018).
journal | preprint ]

Origin of Public Memory B Cell Clones in Fish After Antiviral Vaccination
Magadan S, Jouneau L, Puelma Touzel M, Marillet S, Chara W, Six A, Quillet E, Mora T, Walczak A, Cazals F, Sunyer O, Fillatreau S, Boudinot P.
Front. Immunol. (2018).
journal ]

Invasions of Host-Associated Microbiome Networks
Murall CL, Abbate JL, Puelma Touzel M, Allen-Vercoe E, Alizon S, Froissart R, & McCann K.
Chapter in "Networks of Invasion: Empirical Evidence and Case Studies". Advances in Ecological Research (2017).
journal | pdf ]

Cellular dynamics and stable chaos in balanced networks
PhD Thesis, University of Goettingen (2015).
online library ]

Complete Firing-Rate Response of Neurons with Complex Intrinsic Dynamics
Puelma Touzel M, & Wolf F.
PLoS Computational Biology (2015).
journal ]

Dynamical models of cortical circuits
Wolf F, Engelken R, Puelma-Touzel M, Weidinger JDF, & Neef A.
Current Opinion in Neurobiology (2014).
journal | pdf ]

Optimal bounded-error strategies for projective measurements in nonorthogonal-state discrimination
Touzel MAP, Adamson RBA, & Steinberg AM.
PRA (2007).
journal | preprint ]

Teaching Experience

Advocacy & Service