Russ Salakhutdinov

Ruslan Salakhutdinov is a Canadian machine-learning researcher, the UPMC Professor of Computer Science in the Machine Learning Department at Carnegie Mellon University, and Chief Scientist of Magic, the San Francisco coding-foundation-model lab. He was the first Director of AI Research at Apple from 2016 through 2020 and a doctoral student of Geoffrey Hinton at the University of Toronto. He is a co-author of the foundational deep-learning papers "Reducing the Dimensionality of Data with Neural Networks" (Hinton and Salakhutdinov, Science 2006), "Deep Boltzmann Machines" (Salakhutdinov and Hinton, AISTATS 2009), and "Dropout" (Srivastava et al., JMLR 2014).

At a glance

• Education: BS in Computer Science and Mathematics, High Point University, NC (1998 to 2001), Honors. MSc (2001 to 2003) and PhD (2005 to 2009) in Computer Science, University of Toronto. PhD supervised by Geoffrey Hinton with the thesis Learning Deep Generative Models.
• Current roles: UPMC Professor of Computer Science, Machine Learning Department, Carnegie Mellon University, since February 2016. Chief Scientist at Magic.
• Previous role: Director of AI Research at Apple, November 2016 to January 2020, concurrent with CMU; joined Apple following the 2016 acquisition of Perceptual Machines, which he co-founded in 2015.
• Key contributions: "Reducing the Dimensionality of Data with Neural Networks" (Hinton and Salakhutdinov, Science 2006); "Deep Boltzmann Machines" (Salakhutdinov and Hinton, AISTATS 2009); "Dropout" (Srivastava et al., JMLR 2014); "Show, Attend and Tell" (Xu, Ba et al., ICML 2015). General Chair of ICML 2024, Program Co-Chair of ICML 2019.
• Awards: Sloan Research Fellowship (2013); Microsoft Research Faculty Fellowship (2013); Google Faculty Award (2014); Canada Research Chair in Statistical Machine Learning (2016); Nvidia Pioneers of AI Research (2016); CIFAR Senior Fellow (2011).
• X / Twitter: @rsalakhu
• LinkedIn: russ-salakhutdinov-53a0b610
• Personal site: cs.cmu.edu/~rsalakhu
• Google Scholar: Ruslan Salakhutdinov

Origins

Salakhutdinov was born around 1980 in Tashkent, Uzbekistan, and is of Tatar origin. He completed undergraduate study at High Point University in North Carolina from 1998 to 2001, graduating with an honors double major in Computer Science and Mathematics. He then moved to Toronto, completing the Master of Science in Computer Science at the University of Toronto from 2001 to 2003 under Sam Roweis with the thesis "Optimization Algorithms for Learning". Between the master's and doctoral programs he worked at Canadian Imperial Bank of Commerce in Toronto from 2003 to 2005, before returning to the University of Toronto to begin doctoral study with Geoffrey Hinton in September 2005.

Career

Salakhutdinov's PhD ran from September 2005 to August 2009 under Hinton. The doctoral period coincided with the deep-learning revival the Hinton group helped catalyze and produced the 2006 Science paper "Reducing the Dimensionality of Data with Neural Networks", recurrently cited as a founding artifact of the modern deep-learning era. The 2009 thesis, Learning Deep Generative Models, focused on probabilistic graphical models for unsupervised representation learning, including Restricted Boltzmann Machines and Deep Belief Networks.

After the PhD, he took a postdoctoral position at MIT in the Brain and Cognitive Sciences department and CSAIL from September 2009 through July 2011, working with Joshua Tenenbaum on probabilistic-program approaches to concept learning. The collaboration produced the 2015 Science paper "Human-level concept learning through probabilistic program induction" (Lake, Salakhutdinov, Tenenbaum).

In August 2011, Salakhutdinov returned to the University of Toronto as an Assistant Professor jointly appointed in Computer Science and Statistical Sciences. He was named a Fellow of the Canadian Institute for Advanced Research Neural Computation and Adaptive Perception Program in September 2011, and received the Sloan Research Fellowship and Microsoft Research Faculty Fellowship in 2013. The Toronto period produced the Dropout paper (Srivastava et al., JMLR 2014) and the "Show, Attend and Tell" image-captioning paper (Xu, Ba et al., ICML 2015).

In 2015 Salakhutdinov co-founded Perceptual Machines, a Pittsburgh-based deep-learning startup. Apple acquired the company in 2016, coinciding with his transition to a dual academic-and-industry posture. He moved to Carnegie Mellon University as an Associate Professor in the Machine Learning Department in February 2016, with the Canada Research Chair in Statistical Machine Learning announced the following month. In October 2016 Apple appointed him as its first Director of AI Research, a position he held concurrently with CMU from November 2016 through January 2020. Under his leadership, Apple's AI research group began submitting papers to academic venues, a departure from the company's prior secrecy-first stance.

He returned to a full-time CMU posture in 2020 and was later promoted to UPMC Professor of Computer Science. He served as Program Co-Chair of ICML 2019 and General Chair of ICML 2024. In 2024 he took on the Chief Scientist role at Magic, the San Francisco coding-foundation-model lab co-founded by Eric Steinberger and Sebastian De Ro, concurrent with the CMU professorship. The appointment connects the lab's Long-Term Memory architectural line to his work on deep generative models and large-context representations.

Affiliations

• High Point University: BS in Computer Science and Mathematics (Honors), 1998 to 2001.
• University of Toronto: MSc, 2001 to 2003; PhD, 2005 to 2009, supervised by Geoffrey Hinton.
• Canadian Imperial Bank of Commerce (CIBC): Toronto, 2003 to 2005.
• MIT: Postdoctoral Research Associate, BCS and CSAIL, September 2009 to July 2011.
• University of Toronto, Computer Science and Statistical Sciences: Assistant Professor, August 2011 to January 2016.
• Perceptual Machines: Co-founder, 2015 (acquired by Apple, 2016).
• Carnegie Mellon University, Machine Learning Department: Associate Professor, February 2016; UPMC Professor of Computer Science (named chair).
• Apple: Director of AI Research, November 2016 to January 2020.
• Magic: Chief Scientist, 2024 to present.
• Canadian Institute for Advanced Research: Senior Fellow, Neural Computation and Adaptive Perception Program, September 2011 to present.

Notable contributions

Salakhutdinov's published record concentrates in deep generative models, probabilistic graphical models, representation learning, and multimodal language models. His Google Scholar profile lists more than 250 publications and 200,000-plus citations as of May 2026.

• "Reducing the Dimensionality of Data with Neural Networks" (Hinton and Salakhutdinov, Science 2006). Introduced deep-autoencoder pretraining and is recurrently cited as a founding artifact of the modern deep-learning era. Approximately 25,000 citations.
• "Deep Boltzmann Machines" (Salakhutdinov and Hinton, AISTATS 2009). First full presentation of the multi-layer Boltzmann-machine generative model that anchored the deep-learning generative-model line through the early 2010s.
• "Restricted Boltzmann Machines for Collaborative Filtering" (Salakhutdinov, Mnih, Hinton, ICML 2007). Application of Restricted Boltzmann Machines to the Netflix Prize collaborative-filtering problem.
• "Dropout: A Simple Way to Prevent Neural Networks from Overfitting" (Srivastava, Hinton, Krizhevsky, Sutskever, Salakhutdinov, JMLR 2014). Co-author paper introducing the dropout regularization technique that became a default component of feedforward and convolutional architectures. Approximately 75,000 citations.
• "Show, Attend and Tell: Neural Image Caption Generation with Visual Attention" (Xu, Ba, Kiros, Cho, Courville, Salakhutdinov, Zemel, Bengio, ICML 2015). Co-author paper introducing soft and hard attention mechanisms for image captioning, with Jimmy Ba as second author.
• "Human-level concept learning through probabilistic program induction" (Lake, Salakhutdinov, Tenenbaum, Science 2015). The Bayesian-Program-Learning paper on one-shot character recognition at human-level accuracy.
• "Importance Weighted Autoencoders" (Burda, Grosse, Salakhutdinov, ICLR 2016). Co-authored paper introducing the IWAE training objective.
• Public-talk record. Multiple ICML and NeurIPS keynotes; the University of Pennsylvania GRASP Robotics seminar on Multimodal AI Agents (October 2024); the Kempner Institute seminar on Multimodal AI Agents (November 2024).

Investments and boards

• Magic (AI): Chief Scientist, 2024 to present.
• Apple (AI): Director of AI Research, 2016 to 2020. Joined through the acquisition of Perceptual Machines.
• Perceptual Machines (AI): Co-founder, 2015 (acquired by Apple, 2016).
• Felix Smart (Software): Board Director, 2023 to present.

No other public investor activity on record in AI, semiconductors, datacenters, software, or energy as of May 2026.

Network

Salakhutdinov's longest-running professional relationship is with his doctoral advisor Geoffrey Hinton, with whom he co-authored the 2006 Science paper, the 2009 Deep Boltzmann Machines paper, the 2014 Dropout paper, and an extensive subsequent body of work. The University of Toronto deep-learning lab cohort produced long-running collaborators: Andriy Mnih, his master's advisor Sam Roweis (deceased 2010), Jimmy Ba as former PhD student and frequent co-author, Ryan Kiros on the multimodal-language-model line, and Nitish Srivastava on the Dropout and multimodal-Boltzmann-machine work. The Joshua Tenenbaum collaboration at MIT produced the Bayesian-Program-Learning line through Brenden Lake.

The CMU Machine Learning Department peer cohort includes Tom Mitchell, Eric Xing, Barnabas Poczos, and Manuel Blum. PhD-student alumni include Zhilin Yang, Devendra Singh Chaplot, Yuhuai (Tony) Wu (formerly of xAI), Emilio Parisotto, Zihang Dai, and Manzil Zaheer, several of whom moved to senior roles at frontier labs. The Magic cohort connects the academic record to the commercial coding-foundation-model frontier. Eric Steinberger and Sebastian De Ro are the principal day-to-day collaborators in the Chief Scientist role.

Position in the field

As of May 2026, Salakhutdinov occupies a structurally distinctive position among senior machine-learning researchers through the Hinton-lineage doctoral pedigree, the academic-and-industry dual-track career posture, the long publication record across deep generative models and multimodal-language-models, and the senior leadership of two major ICML conferences (Program Co-Chair 2019, General Chair 2024). The 2006 Science paper, the Deep Boltzmann Machines paper, and the Dropout co-authorship anchor a founding-generation positioning alongside the Hinton, Bengio, and LeCun cohort.

The Apple period from 2016 to 2020 placed Salakhutdinov among a small set of senior academics who established formal industry-research leadership without leaving university faculty positions, alongside peers like Yann LeCun at Meta. Under his leadership Apple's research group began publishing in peer-reviewed academic venues. The 2024 Magic Chief Scientist appointment connects the academic and industry strands to a specific commercial coding-foundation-model thesis: Magic's Long-Term Memory line and the August 2024 LTM-2-mini announcement draw on long-context representation-learning research aligned with the multimodal and generative-model work at his CMU group.

Outlook

Open questions over the next 6 to 18 months:

• Magic technical-paper cadence. Whether Magic publishes additional research papers on the LTM architecture under Salakhutdinov's research leadership, and whether the publications connect to the deep-generative-model and multimodal-representation-learning lines from his CMU group.
• CMU group research direction. Whether the CMU Machine Learning Department group continues the multimodal-AI-agents and VisualWebArena research line, and whether new directions emerge from the Magic cross-pollination.
• Magic commercial product. Whether Magic launches a broadly available autonomous-coding product comparable to Cursor or GitHub Copilot, and the role of the Chief Scientist research function in the product development.
• CMU-and-Magic balance. Whether the dual-track posture continues at the current cadence or shifts toward one institution, given the historical pattern of senior academics at industry labs eventually moving to single-institution positions.
• Public-talk cadence. Whether the post-ICML-2024-General-Chair schedule continues the moderate seminar pace at the Kempner Institute, GRASP Robotics, and similar venues.

Sources

• Ruslan Salakhutdinov on Wikipedia. Wikipedia entry recording education, career, awards, and publications.
• Russ Salakhutdinov personal site. Personal site at Carnegie Mellon listing publications, teaching, students, and biographical information.
• Russ Salakhutdinov CV. Curriculum vitae listing complete education, professional career, awards, grants, and publication record.
• Russ Salakhutdinov on Google Scholar. Citation metrics and complete publication listing.
• Russ Salakhutdinov on X. The @rsalakhu X account, the primary public channel for his commentary.
• Russ Salakhutdinov on LinkedIn. Public LinkedIn profile.
• Ruslan Salakhutdinov Joins the Machine Learning Department. Carnegie Mellon Machine Learning Department announcement of the February 2016 Associate Professor appointment.
• Apple hires CMU professor as director of AI research to smarten up Siri. TechCrunch coverage of the October 2016 Apple Director of AI Research appointment.
• Reducing the Dimensionality of Data with Neural Networks. The 2006 Science paper introducing deep-autoencoder pretraining.
• Deep Boltzmann Machines. The 2009 AISTATS paper introducing the Deep Boltzmann Machine model.
• Dropout: A Simple Way to Prevent Neural Networks from Overfitting. The 2014 JMLR paper introducing the dropout regularization technique.
• Show, Attend and Tell: Neural Image Caption Generation with Visual Attention. The 2015 ICML paper introducing soft and hard attention mechanisms for image captioning.
• Human-level concept learning through probabilistic program induction. The 2015 Science paper presenting Bayesian Program Learning at human-level accuracy on character recognition.
• Magic blog: 100M Token Context Windows. August 2024 LTM-2-mini announcement from Magic.
• Fall 2024 GRASP on Robotics: Ruslan Salakhutdinov. University of Pennsylvania GRASP seminar abstract and bio (October 2024).
• Multimodal AI Agents - Kempner Institute. Harvard Kempner Institute seminar abstract and bio (November 2024).
• Ruslan Salakhutdinov - GenAI Summit 2025. UC San Diego GenAI Summit 2025 speaker bio.
• Feature image: Russ Salakhutdinov portrait, Wikipedia entry on Russ Salakhutdinov, CC BY 2.0, photographer Steve Jurvetson (cropped from "Deep Thinkers on Deep Learning", October 2016).