Last reviewed
Jun 9, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v3 · 4,507 words
Add missing citations, update stale details, or suggest a clearer explanation.
Mila, formally styled Mila - Quebec Artificial Intelligence Institute (in French, Mila - Institut québécois d'intelligence artificielle), is a non-profit research institute headquartered in the Mile-Ex district of Montreal, Quebec, Canada. By head count of affiliated professors it is the largest academic deep learning research community in the world, reporting more than 140 core and associate faculty members and a broader community of roughly 1,400 students, postdoctoral fellows, technology specialists, and industry partners as of 2025. It was founded in 1993 by yoshua bengio as a small machine learning lab at the Université de Montréal and grew, in stages, into a federated institute that today brings together researchers from the Université de Montréal, McGill University, HEC Montréal, Polytechnique Montréal, and several other Quebec universities.[1][2]
Mila is one of three national AI institutes anchoring the Canadian government's Pan-Canadian Artificial Intelligence Strategy, alongside the Vector Institute in Toronto and the Alberta Machine Intelligence Institute (Amii) in Edmonton. Its scientific output spans foundational work on deep neural networks and word embeddings, contributions to generative modelling and attention mechanisms in neural machine translation, research in reinforcement learning, and a visible programme of work on AI safety, governance, and responsible development. Its founder, Bengio, shared the 2018 ACM A.M. turing award with Geoffrey Hinton and Yann LeCun for foundational contributions to deep learning; in 2025 he stepped back from the role of scientific director to focus on AI safety initiatives, including the chairmanship of the International AI Safety Report and the founding of the non-profit LawZero.[3][4][5]
This article covers Mila's founding and naming history, scientific and governance structure, faculty and alumni, principal research areas, and leading role in international AI safety discussions from 2023 to 2026.
Mila traces its origins to 1993, when Yoshua Bengio joined the Department of Computer Science and Operations Research (DIRO) at the Université de Montréal as an assistant professor and established a small research group. The group was initially called the Laboratoire d'informatique des systèmes adaptatifs (LISA), the Laboratory for Computer Science of Adaptive Systems. LISA's research programme sat at the intersection of artificial intelligence, statistical inference, and adaptive computation: the lab worked on neural networks, statistical language models, and learning algorithms throughout the 1990s and 2000s, when those topics were marginal in mainstream machine learning research.[2][6]
Through the late 2000s and 2010s, LISA became one of the most productive deep learning groups in the world. Its early work on probabilistic neural language models, word embeddings, deep architectures, denoising autoencoders, and stochastic gradient methods contributed directly to the deep learning revolution.[1][6] LISA also developed Theano, an open-source numerical computation library that was, for several years, one of the dominant deep learning frameworks; later frameworks, including early versions of TensorFlow and PyTorch, drew on concepts and contributors from the project.[2]
Around 2016 and 2017, LISA was rebranded internally as MILA, an acronym for the Montreal Institute for Learning Algorithms, reflecting the lab's growth into a multi-institutional community. In 2017, alongside the announcement of the Pan-Canadian Artificial Intelligence Strategy, MILA was reorganised as a partnership between the Université de Montréal and McGill University, with close collaboration from HEC Montréal and Polytechnique Montréal. The institute was incorporated as a non-profit in 2018 under the new name Mila - Quebec Artificial Intelligence Institute. The lowercase form "Mila" became the official styling, replacing the all-caps acronym, and the institute's English designation now references "Quebec Artificial Intelligence Institute" rather than the older "Montreal Institute for Learning Algorithms".[2][7]
In January 2019, Mila inaugurated its current premises at the O Mile-Ex complex at 6666 Saint-Urbain Street, in the Mile-Ex neighbourhood of Montreal. The opening took place with Quebec Minister of Economy and Innovation Pierre Fitzgibbon present, alongside more than 200 partners from across the Quebec AI ecosystem. The new building offered more than 90,000 square feet of recycled industrial space and was designed to house Mila's professors, students, technology development team, a startup accelerator called Espace CDPQ | Axe IA, and a dozen corporate research labs co-located on site.[7][8]
For most of Mila's history, the institute's public identity was closely tied to its founder, Yoshua Bengio. Born in Paris in 1964 and raised in Montreal, Bengio completed his PhD at McGill University in 1991 before postdoctoral work at MIT and AT&T Bell Labs. He returned to Montreal as a faculty member at the Université de Montréal in 1993 and began the work that would become Mila.[3]
Bengio served as scientific director of Mila from its founding until 2025, leading research direction and acting as its principal public representative. He was awarded the 2018 ACM A.M. Turing Award jointly with Geoffrey Hinton and Yann LeCun for "conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing", and is grouped with Hinton and LeCun as the "godfathers" of deep learning. Bengio is the most-cited computer scientist in the world by both total citations and h-index, and in November 2025 he became the first AI researcher to surpass one million Google Scholar citations.[3][4]
On 28 March 2025, Mila announced a formal transition in its scientific leadership. Bengio assumed a newly created role of Founder and Scientific Advisor, while Laurent Charlin, a professor at HEC Montréal and a long-time Mila core member, was appointed Interim Scientific Director. The institute stated that the transition would allow Bengio to devote more time to his research on AI safety and to international governance initiatives, and that a recruitment process had been launched for a permanent successor.[9] On 2 September 2025, Mila announced that Hugo Larochelle had been appointed as the new Scientific Director, succeeding Charlin and reporting effective immediately. Larochelle, an adjunct professor at the Université de Montréal and a former head of Google's Montreal AI research lab, had originally trained as a doctoral student under Bengio at the Université de Montréal in the 2000s; Bengio publicly described their relationship as a collaboration "spanning more than 20 years".[10][11]
Mila is a non-profit organisation rather than a single university department, and its faculty hold their primary academic appointments at partner universities. Four institutions are identified as Mila's founding university partners:
In addition, Mila hosts professors from Université Laval in Quebec City, Université de Sherbrooke, the École de technologie supérieure (ÉTS) in Montreal, and Concordia University. Together these eight Quebec post-secondary institutions form the pool of universities from which Mila's academic members are drawn.[1][2]
The institute is governed by a board of directors, with Pierre Boivin as Board President and Valérie Pisano as chief executive officer. Scientific direction is provided by the Scientific Director, supported by Associate Scientific Co-Directors, a Scientific Council, and strategic priority leads. Under Phase 2 of the Pan-Canadian AI Strategy, announced in Budget 2021, the federal government committed up to CAD 60 million for the three national institutes (Mila, Amii, and Vector), with each eligible for up to CAD 20 million over five years between 2021 and 2026. The Government of Quebec separately committed CAD 36 million in 2024, on top of an earlier CAD 21 million over three years for "socially beneficial AI" research.[12][13]
By 2025, Mila reported more than 140 affiliated professors and a broader community of around 1,400 researchers including PhD students, master's students, postdoctoral fellows, and technology specialists, making it the largest concentration of academic deep learning researchers in any single institute by faculty count.[1][14] Several Mila professors are also Canada CIFAR AI Chairs, the flagship recruitment programme of the Pan-Canadian AI Strategy that supports more than 100 researchers across the three national institutes.[15]
Beyond Bengio and Larochelle, prominent affiliated researchers include:
Several researchers trained at Mila (or its predecessor LISA/MILA) have gone on to major roles in the broader AI ecosystem.
Mila's research portfolio has evolved over time, but its core scientific identity rests on deep learning and its applications. The institute groups its activities into broadly the following themes.
Mila's faculty have made foundational contributions to neural network theory and practice over more than two decades. Early LISA work on probabilistic neural language models (Bengio et al., 2003) helped lay the groundwork for word embeddings and modern language modelling. Later contributions include training of deep networks, regularisation, optimisation, denoising autoencoders, and energy-based models.[1][2][6] The textbook Deep Learning by Goodfellow, Bengio, and Courville (MIT Press, 2016) summarises much of this work and is one of the most cited textbooks in the field.[16]
Mila is the original home of the generative adversarial network: the 2014 paper "Generative Adversarial Nets" by Ian Goodfellow and collaborators, including Bengio, was conducted while Goodfellow was a PhD student at the Université de Montréal under Bengio's supervision. Mila researchers have continued to work extensively on generative models, including variational autoencoders, normalising flows, score-based and diffusion models, and large-scale generative models for text, images, molecules, and proteins.[19]
The first attention mechanism for neural machine translation, often called "Bahdanau attention", was introduced in a 2014 paper by Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio, all then affiliated with Mila. This work, together with the Cho et al. 2014 paper introducing the gated recurrent unit and an early version of the encoder-decoder architecture, is widely credited as a direct intellectual precursor to the Transformer architecture introduced by Google researchers in 2017.[20]
Reinforcement learning is a long-standing area of strength at McGill's RL-Lab, and through it at Mila. Researchers including Doina Precup and Joelle Pineau have contributed to options frameworks, temporal abstraction, off-policy evaluation, and applications to dialogue systems and healthcare. Several Mila faculty hold joint positions with Google DeepMind's Montreal office, reflecting the close coupling between Mila and DeepMind in reinforcement learning research.[17]
A significant fraction of Mila's research output is in applications of machine learning to scientific and societal problems. Mila has run dedicated programmes on AI for climate change (including the founding of organisations such as Climate Change AI by Mila-affiliated researchers), AI for drug discovery and molecular design (including collaborations with Valence Discovery and Recursion), and AI for healthcare. The institute also hosted prominent COVID-19 modelling efforts during the 2020 to 2022 period.[1]
From the late 2010s, Mila has built a substantial programme in AI safety, alignment, and responsible AI development. This work overlaps with international research on superalignment, mechanistic interpretability, and risk assessment for advanced general-purpose AI. The institute's "AI for Humanity" cluster supports work on fairness, bias, ethics, governance, and human rights in AI, and it co-developed the Montreal Declaration for a Responsible Development of Artificial Intelligence, first announced in November 2017 and built around ten core values including well-being, autonomy, privacy, solidarity, democratic participation, equity, diversity and inclusion, prudence, responsibility, and sustainable development.[21]
Beginning around 2022, Yoshua Bengio shifted a substantial portion of his personal research effort toward AI safety, and Mila increasingly became identified with that work. Three interlinked initiatives stand out: Bengio's endorsement of the 2023 "pause" open letter from the Future of Life Institute, his role as chair of the United Kingdom-commissioned International AI Safety Report, and his founding of the safety non-profit LawZero in 2025.
In March 2023, the Future of Life Institute published an open letter titled "Pause Giant AI Experiments", calling on all AI labs to pause for at least six months the training of AI systems "more powerful than GPT-4". The letter gathered more than 30,000 signatures, including from Yoshua Bengio, the computer scientist Stuart Russell, Elon Musk, Steve Wozniak, the historian Yuval Noah Harari, and others. Bengio explicitly endorsed the six-month pause in a press conference, citing concerns about the rate of capability progress and the difficulty of supervising increasingly autonomous systems. The letter's text had been developed in consultation with experts including Bengio and Russell.[22]
The pause letter is widely regarded as the moment at which mainstream debate over catastrophic and existential risks from frontier AI became visible to a broad audience. Bengio's signature placed Mila's founder firmly within the international community of researchers, including those at organisations such as far ai and conjecture, advocating for slower, more careful frontier AI development.
In November 2023, the United Kingdom hosted the first AI Safety Summit at Bletchley Park, whose communiqué committed the participating 30 nations to support an international, scientifically grounded synthesis of evidence on the safety of advanced AI, modelled in part on the Intergovernmental Panel on Climate Change. Bengio was named chair of the report. An interim version, originally titled the International Scientific Report on the Safety of Advanced AI, was published in May 2024 ahead of the AI Seoul Summit, and the first full International AI Safety Report 2025 was published on 29 January 2025 ahead of the France AI Action Summit in Paris. The report was prepared by 96 listed AI experts under Bengio's chairmanship, supported by an Expert Advisory Panel drawn from 30 nations plus the United Nations, OECD, and European Union.[23]
The 2025 report assessed capabilities, risks, and mitigations for general-purpose AI but explicitly did not make policy recommendations; its purpose was to provide a shared international scientific baseline for policymakers. A first "Key Update" appeared in October 2025, focused on rapid capability advances in reasoning language models and new evidence of difficulties in monitoring frontier systems. A second update followed in November 2025, and the second full annual edition was published on 3 February 2026 ahead of the India AI Impact Summit.[23][24] While the report is an intergovernmental product rather than a Mila product, its association with Bengio cemented Mila's reputation as a leading voice in AI safety policy alongside bodies such as the UK AI Security Institute and the US AI Safety Institute.
On 3 June 2025, Bengio publicly launched LawZero, a non-profit AI safety research organisation based in Montreal. The launch was accompanied by reporting in The Globe and Mail, Reuters, Le Devoir, The Times of London, TechCrunch, Axios, and other major outlets. LawZero's stated mission is to advance "research and creating technical solutions that enable safe-by-design AI systems", with a particular focus on what Bengio calls "Scientist AI": a non-agentic class of system designed to "understand the world rather than act in it", to issue calibrated Bayesian probabilistic statements rather than to take open-ended actions in the world, and to serve as a safety guardrail and oversight tool for other AI systems.[25][26]
Initial reporting by TechCrunch and others put LawZero's launch funding at approximately USD 30 million in philanthropic contributions from a coalition including the Future of Life Institute, Schmidt Sciences (a successor to the philanthropy of former Google CEO Eric Schmidt), Skype founding engineer Jaan Tallinn, Open Philanthropy, and the Silicon Valley Community Foundation. LawZero is staffed by more than 15 researchers at launch, with Bengio serving as President and Scientific Director. The organisation is structured as a non-profit to insulate it from commercial and government pressures.[25][26]
LawZero is a legally separate organisation from Mila, but it is led by Mila's founder and is based in Montreal. Bengio has indicated that his Mila-affiliated research group continues to work on safety-relevant deep learning problems, while LawZero focuses on translating that research into deployable safe-by-design systems. Together with overseas peers such as far ai, conjecture, and the work of aria uk on safeguarded AI under Programme Director David "davidad" Dalrymple, LawZero forms part of a small but growing ecosystem of non-profit and quasi-governmental institutions devoted to safety-focused AI research outside the major commercial frontier labs.
Mila's headquarters at 6666 Saint-Urbain Street, in the O Mile-Ex building, occupies more than 90,000 square feet of converted industrial space and houses Mila's professors, students, technology development team, on-site corporate research labs, and an early-stage startup accelerator. The Mile-Ex neighbourhood, between Mile End and Parc-Extension, has been developed by the City of Montreal and provincial agencies as an AI cluster, with companies including Element AI (acquired by ServiceNow in 2020), DeepMind's Montreal office, Microsoft Research Montreal, Google's Montreal AI lab, and IVADO Labs located in the immediate vicinity.[7][8]
Funding sources for Mila include:
In 2024 and 2025, Mila publicly explored launching a dedicated venture capital fund, in collaboration with Inovia Capital and other Montreal-based investors, with reported ambitions of raising up to CAD 100 million to commercialise spin-out companies arising from Mila research.[27]
In addition to scientific research, Mila operates extensive training and engagement programmes, including the long-running CIFAR Deep Learning and Reinforcement Learning Summer School, annual research conferences, and continuing education. Through its AI for Humanity cluster it convenes the Mila AI Policy Conference and, since 2025, the Mila AI Policy Fellowship. The inaugural fellowship cohort, announced in August 2025, brought together nine fellows pursuing six-month projects across seven thematic areas including AI safety and governance, AI adoption, media and democracy, substantive equality, climate, Indigenous rights, and public sector innovation.[28]
Mila also helped coordinate the Montreal Declaration for a Responsible Development of Artificial Intelligence, launched on 3 November 2017 by the Université de Montréal alongside Mila. The declaration is built on ten core values: well-being, respect for autonomy, protection of privacy and intimacy, solidarity, democratic participation, equity, diversity and inclusion, prudence, responsibility, and sustainable development. Mila uses it as a reference document for institutional ethics.[21]
Mila is one of three national AI institutes funded under the Pan-Canadian AI Strategy. It is useful to distinguish it carefully from the other two:
Internationally, Mila is sometimes compared to AI research centres outside Canada, including the Center for Human-Compatible AI (CHAI) at the University of California, Berkeley, founded in 2016 by Stuart Russell; the Stanford Institute for Human-Centered Artificial Intelligence (Stanford HAI), founded in 2019; and smaller university and non-profit centres in the United Kingdom and continental Europe. Mila differs in scale (it is substantially larger by faculty count than CHAI and uses a different organisational model from Stanford HAI, which is a Stanford-internal institute) and in its tight integration with industry partners and government strategy.
The years 2024 to 2026 have been a period of substantial change at Mila, driven by the maturation of generative AI and by Bengio's pivot toward AI safety.
The institute has continued to grow rapidly: Mila announced 20 new faculty in 2024 and a further 13 in 2025, reaching more than 140 affiliated professors and roughly 1,400 affiliated researchers overall.[14][29] Its scientific leadership has transitioned for the first time in over three decades, with Bengio's move to Founder and Scientific Advisor in March 2025, Charlin's interim service, and Larochelle's appointment as permanent Scientific Director on 2 September 2025.[9][10][11]
AI safety has become institutionally central. Beyond Bengio's role on the International AI Safety Report and the launch of LawZero, Mila has expanded its AI Governance, Policy and Inclusion programme and launched the Mila AI Policy Fellowship.[28] The institute's 2024 to 2025 Impact Report explicitly identifies responsible AI, AI safety, and AI policy as strategic priorities.[14] Mila has also deepened its commercial partnerships and explored mechanisms for commercialising research, including the prospective Inovia-affiliated venture fund reported in late 2024 and 2025.[27]
The institute's research portfolio has broadened beyond core deep learning into multi-agent reasoning, foundation models for science (chemistry, biology, materials), and safety-focused work on interpretability, evaluations, and verification. As of 2026, Mila remains the world's largest concentration of academic deep learning researchers by faculty count and, through its founder's external roles, one of the most prominent institutional voices in the international debate over the governance of advanced AI.[14][23][25]