Hugo Larochelle

wiki
, , , ,
1

Canadian deep learning researcher and Scientific Director of Mila – Québec AI Institute, known for foundational contributions to neural autoregressive models, denoising autoencoders, and zero-shot learning.

Profile
Born 1981, Canada
Nationality Canadian
Current Institution Mila – Québec AI Institute; Université de Montréal (Adjunct); McGill University (Adjunct)
Research Areas Deep Learning, Generative Models, Representation Learning, Zero-Shot Learning
Doctoral Advisor Yoshua Bengio
Doctoral Thesis Études de techniques d’apprentissage non-supervisé pour l’amélioration de l’entraînement supervisé de modèles connexionnistes (2009)
Personal Website cs.toronto.edu/~larocheh
X / Twitter @hugo_larochelle
GitHub larocheh

Overview

Hugo Larochelle is a Canadian computer scientist whose doctoral and postdoctoral work placed him at the heart of the Montreal deep learning ecosystem during its formative years. His research produced several techniques now embedded in modern AI systems, including denoising autoencoders, neural autoregressive distribution models, and an early formalisation of zero-shot learning. After a decade spanning academia, a startup acquisition, and leadership of Google’s Montreal AI lab, he was appointed Scientific Director of Mila, the Québec AI Institute, in September 2025, succeeding Laurent Charlin in the role — effectively stepping into the institutional seat long associated with his own doctoral supervisor, Yoshua Bengio. He holds adjunct professorships at both the Université de Montréal and McGill University, and is a Canada CIFAR AI Chair.

Early Life & Education

Hugo Larochelle was born in 1981 in Canada and pursued his entire university education at the Université de Montréal. He earned a BSc in Mathematics and Computer Science in 2004, then remained at the same institution for doctoral study.

His PhD in Computer Science was completed in 2009 under the supervision of Yoshua Bengio. The thesis, titled Études de techniques d’apprentissage non-supervisé pour l’amélioration de l’entraînement supervisé de modèles connexionnistes (Studies of unsupervised learning techniques for improving the supervised training of connectionist models), explored pretraining strategies for deep networks at a time when the field had yet to gain mainstream acceptance. As Larochelle later recalled, early work involving artificial neural networks was met with considerable scepticism: the first paper he worked on was met with disinterest because it involved artificial neural nets, a concept that had yet to gain much traction.

From 2009 to 2011, he completed postdoctoral research at the University of Toronto under the supervision of Geoffrey Hinton — the second of the two figures now widely called “Godfathers of AI” who shaped his early career.

Career

Université de Sherbrooke (2011–2016)

Larochelle joined the Département d’informatique at the Université de Sherbrooke as an Assistant Professor in 2011. During this period he continued publishing prolifically on generative models and representation learning while teaching courses that would later become the basis for his widely circulated online lecture series. It was also during this time that he began the applied work that would lead to a startup.

Whetlab / Twitter (2014–2016)

While at Sherbrooke, Larochelle co-founded Whetlab alongside Jasper Snoek, Kevin Swersky, Ryan P. Adams, and Alex Wiltschko. The company focused on practical Bayesian optimisation techniques for automating machine learning hyperparameter tuning. Whetlab was acquired by Twitter in June 2015. Following the acquisition, Larochelle moved to Twitter as a Research Scientist in the Twitter Cortex group.

Google Brain Montreal (2016–2023)

In late 2016, Larochelle left Twitter to establish and lead a new Google Brain research group in Montreal — Google’s first dedicated AI research lab in Canada. He was recruited to lead Google’s AI research lab in Montreal, which is now integrated into Google DeepMind. Under his direction, the lab grew into one of Google’s most productive research outposts, working on areas from reinforcement learning and generative modelling to applications in environmental science and code generation.

Google DeepMind Montreal (2023–2025)

In April 2023, Google Brain merged with DeepMind to form Google DeepMind. Larochelle worked as a Principal Scientist within the unified Google DeepMind organisation in Montreal until April 2025.

Mila – Québec AI Institute (2025–present)

In September 2025, Mila announced the appointment of Hugo Larochelle to the position of Scientific Director. He officially assumed his duties on September 2, succeeding Laurent Charlin, who had been acting in that capacity since Yoshua Bengio transitioned to the role of founder and Scientific Advisor of Mila. In this capacity, he leads an institute with a community of over 1,500 researchers and 150 industrial partners. His stated vision emphasises interdisciplinary research he calls “AI and x,” connecting machine learning with fields such as ecology, health, and environmental science.

Key Contributions

  • Denoising Autoencoders (DAE) — Co-authored the foundational 2008 ICML paper and its 2010 JMLR extension with Pascal Vincent, Yoshua Bengio, and others. The work established corrupted-input reconstruction as a scalable paradigm for learning meaningful representations from large quantities of unlabelled data, and remains a cornerstone reference in self-supervised learning.

  • Zero-Data Learning of New Tasks — A 2008 AAAI paper with Dumitru Erhan and Yoshua Bengio that formalised the concept of generalising to categories for which no labelled training examples exist. This work introduced the now-standard concept of zero-shot learning that underpins a large body of modern transfer and few-shot learning research.

  • Neural Autoregressive Distribution Estimator (NADE) — Introduced at AISTATS 2011 with Iain Murray, winning a Notable Paper Award. NADE offered a tractable and scalable approach to modelling multivariate distributions that helped establish neural autoregressive modelling as a practical paradigm. Through models such as NADE and MADE, Larochelle helped popularise the neural autoregressive modelling paradigm now omnipresent in generative AI.

  • Greedy Layer-Wise Training of Deep Networks — Co-authored with Yoshua Bengio and colleagues at NIPS 2007, one of the seminal papers demonstrating that deep networks could be effectively trained via layer-wise unsupervised pretraining.

  • Exploring Strategies for Training Deep Neural Networks — A comprehensive 2009 JMLR study providing systematic empirical guidance on training deep architectures, widely cited in the subsequent decade of deep learning development.

  • Transactions on Machine Learning Research (TMLR) — Larochelle co-founded TMLR alongside Raia Hadsell and Kyunghyun Cho; the journal launched in 2022 under the umbrella of the Journal of Machine Learning Research. He served as founding Editor-in-Chief until the end of 2025, shaping a publication model designed to provide faster, more open peer review for the ML community.

  • Neural Networks Online Course — Created a comprehensive video lecture series on neural networks and deep learning based on courses delivered at the Université de Sherbrooke, made freely available on YouTube and widely used by learners and educators worldwide.

  • TechAide Conference — Founded the annual TechAide event, mobilising Montreal’s technology community to raise funds for Centraide to support its mission of combating poverty and social exclusion.

Awards & Recognition

  • AISTATS Notable Paper Award (2011) — Awarded to “The Neural Autoregressive Distribution Estimator” co-authored with Iain Murray.
  • Canada CIFAR AI Chair — A competitive research chair recognising leading AI researchers in Canada, held in conjunction with the CIFAR Learning in Machines & Brains programme.
  • Turing Award lineage — Trained directly under both Yoshua Bengio (Turing Award 2018) and Geoffrey Hinton (Turing Award 2018), an unusual academic pedigree that situates Larochelle at the centre of the Montreal deep learning school.

Key Relationships

  • Yoshua Bengio — Doctoral supervisor at the Université de Montréal; Turing Award laureate and founder of Mila, whose Scientific Director role Larochelle now holds.
  • Geoffrey Hinton — Postdoctoral supervisor at the University of Toronto; Turing Award laureate and one of the principal figures of the deep learning revolution.
  • Pascal Vincent — Frequent early collaborator; co-author on the denoising autoencoders papers that became landmarks of self-supervised learning.
  • Iain Murray — Co-author of the NADE paper (AISTATS 2011 Notable Paper); long-standing collaborator on generative modelling.
  • Jasper Snoek & Kevin Swersky — Co-founders of Whetlab; collaborators on Bayesian hyperparameter optimisation research that underpinned the startup.
  • Raia Hadsell & Kyunghyun Cho — Co-founders of TMLR; the three collaborated to design and launch the journal as a new peer-review model for ML research.
  • Aaron Courville — Mila colleague and long-term collaborator on deep learning and reinforcement learning research.
  • Laurent Charlin — Immediate predecessor as interim Scientific Director of Mila; associate professor at HEC Montréal.

Personal Style

Larochelle’s research style is characterised by careful empirical rigour and a preference for tractable, principled probabilistic models over purely engineering-driven approaches — a sensibility that reflects his training in the Bengio school. In public discourse he is measured and collegial, rarely polemical, and tends to frame the importance of academic AI research in terms of its capacity to pursue questions that industrial labs may deprioritise. Since taking on the Scientific Director role at Mila, he has articulated a vision centred on interdisciplinary application — “AI and x” — and has been outspoken about the continued relevance of open, university-rooted research in an era of large corporate laboratories. As a philanthropist, he and his wife have made substantial personal financial commitments to fund graduate scholarships in AI applied to environmental problems, reflecting a consistent alignment between his stated values and his professional choices.

References