Ali Rahimi is a computer scientist and researcher at Google Research, recognized for contributions to machine learning, optimization, and signal processing. He gained widespread attention in the machine learning community for his 2017 keynote at the Conference on Neural Information Processing Systems (NIPS), where he criticized the field for lacking rigorous theoretical foundations, comparing certain practices to alchemy. His work spans algorithmic innovations and practical applications, and he has held roles in both industry and academia.
Rahimi's research interests include convex optimization, randomized algorithms, and their applications in machine learning. He has co-authored numerous papers, and his work has influenced areas such as distributed computing and large-scale data processing. His career includes positions at Intel Labs and Google, where he has worked on projects involving mobile sensing and machine learning infrastructure.
Early Life and Education
Rahimi received his Bachelor of Science degree in Computer Science and Engineering from the massachusetts-institute-of-technology (MIT) in 1996. He then pursued graduate studies at the university-of-california-berkeley, where he completed a Master of Science degree in 200ù, followed by a Ph.D. in Electrical Engineering and Computer Sciences in 2006. His doctoral advisors were Daphne Koller and trevor-darrell, and his dissertation focused on practical approaches to sensor network localization and calibration.
During his time at Berkeley, Rahimi worked on the Smart Dust project, developing miniaturized sensor nodes. This work led to several publications on energy-efficient sensing and collaborative signal processing. He also collaborated with researchers at the BAIR (Berkeley AI Research) lab, where he explored probabilistic graphical models for multimodal data fusion.
Career at Intel Labs
After completing his doctorate, Rahimi joined Intel Labs in Santa Clara, California, in 2006 as a research scientist. At Intel, he led projects on context-aware computing and developed algorithms for human activity recognition using wearable sensors. His team demonstrated early prototypes of wrist-mounted devices that could infer user gestures, which informed later smartwatch technology.
In 2007, Rahimi co-authored a paper titled "Probabilistic Data Fusion for Robust Activity Recognition" with colleagues from Intel and the University of Toronto, which introduced a Bayesian framework for combining accelerometer and microphone data. This work was later cited in research on mobile health monitoring and was featured in several IEEE conferences.
Move to Google and Early Projects
In 2009, Rahimi joined Google as a staff research scientist in the New York City office. He initially worked on the Gmail priority inbox system, developing classification algorithms that ranked emails by importance. This project, launched in 2010, used a gradient-boosted decision tree model trained on user-labeled examples.
Rahimi also contributed to the Android operating system team, where he worked on low-power sensor fusion techniques. He developed an efficient implementation of the Kalman filter that became part of the Android sensor stack. This work was described in a 2012 blog post by the Android team and influenced later versions of the platform's motion processing.
The NIPS 2017 Talk and Aftermath
Rahimi became widely known for his invited talk at NIPS 2017 held in Long Beach, California, on December 5, 2017. In the talk, titled "Machine Learning: The Alchemists," he argued that many widely used machine learning techniques, such as Deep learning models, lack formal guarantees and are used in ways that resemble medieval alchemy. He presented examples where neural networks fail unexpectedly, like adversarial examples, and called for more rigorous theory.
The talk sparked a debate that was covered by MIT Technology Review and other outlets. Responses came from researchers such as Christopher Bishop and Yann LeCun, who argued for pragmatism. Rahimi and Chris Bishop later co-authored a follow-up piece in the NIPS 2017 workshop proceedings, clarifying their positions and suggesting that the field should develop empirical methods with theoretical backing.
Research Areas and Contributions
Beyond the talk, Rahimi's research includes work on random features and kernel methods. He co-authored the 2007 paper "Random Features for Large-Scale Kernel Machines" with ben-recht, a contribution that introduced the concept of approximating shift-invariant kernels using random projections. This paper has been cited over 5,000 times and is a basis for scalable support vector machines.
In the 2010s, Rahimi worked on distributed optimization and data centers. He published a 201orp paper on asynchronous parallel coordinate descent with Michael I. Jordan and others, which analyzed convergence rates for systems with delayed updates. This work was used by Amazon Web Services in its early machine learning services and by other cloud providers.
Rahimi has also investigated signal processing for autonomous vehicles. He collaborated with Waymo engineers on sensor calibration and fusion for lidar and camera data. In 2019, he gave a talk at the Google Cloud Next conference about using probabilistic models to improve object detection in low-visibility conditions.
Awards and Recognition
In 2014, Rahimi received a Google Research Award for his work on privacy-preserving machine learning. He was named a Distinguished Scientist by the ACM in 2018, recognizing his contributions to applied machine learning. In 2020, he received a Test of Time Award at the ICML for the 2007 random features paper, an honor given for papers that have influenced subsequent research and practice.
Rahimi has served on the program committees of major conferences, including NeurIPS, ICML, and CVPR. He has also mentored several Ph.D. students through Google's internship program, some of whom have gone on to academic positions.
Current Work at Google
As of 2024, Rahimi continues to work at Google Research in the Zurich office. He is part of the human-computer interaction team, focusing on machine learning for mobile health applications. His recent projects include a neural network that predicts hypoglycemia from continuous glucose monitor data, developed in collaboration with the Stanford Diabetes Research Center.
He has also been involved in internal Google efforts to improve the interpretability of Large language modelsaine. In 2023, he co-authored a paper on attribution methods for transformer models, examining how attention layers contribute to predictions. This work was presented at the acl annual meeting and has been cited in studies on model auditing.
Publications and Selected Works
Rahimi has authored or co-authored over 50 peer-reviewed papers. Notable examples include the 2003 paper "Simultaneous Calibration and Tracking with a Network of Range Sensors" (with Daphne Koller and trevor-darrell), which appeared in the Proceedings of the IEEE. His 2004 paper "A Method for Measuring the Distribution of Perpendicular Winds" was published in the Journal of Atmospheric and Oceanic Technology and described a Bayesian approach to wind estimation using sensor arrays.
The 2007 random features paper remains his most cited work since 2020. He also published the 2010 paper "On the Difficulty of Learning with Weak Indicators" with osman-ozturk and philip-liang, which analyzed the challenge of training classifiers with noisy labels, a topic that became central to later research in robust learning.
Public Lectures and Writing
Rahimi has given public lectures beyond NIPS. In 2018, he presented "The Next Decade of Machine Learning" at the Amazon AI conference in Berlin, where he argued for more cautious deployment of AI in safety-critical domains. He has also written opinion pieces for the Open Source advocacy blog, discussing reproducibility in ML, and has been a guest on podcasts such as the Talking Machines show.
In 2021, Rahimi co-taught a course at Stanford on machine learning for embedded systems. The course material included his lecture notes on efficient inference and quantization, which were later adapted for use in industry training programs.
Current Work and Interests
As of 2025, Rahimi continues to work at Google DeepMind on formal methods for machine learning. He has been involved in efforts to create provable guarantees for transformer models, related to transformer architectures. He collaborates with researchers at OpenAI and Anthropic on safety evaluations, though he does not hold an official position there.
Rahimi has also spoken about the need for better benchmarking in AI. In 2023, he organized a workshop on reproducibility at the International Conference on Learning Representations (ICLR), which led to the adoption of new reporting guidelines for several publication venues.
Public Speaking and Writing
Rahimi is an active public speaker. He gave a TEDx talk in 2018 in Zurich titled "The Alchemy of AI," which has been viewed over 200,000 times. He maintains a blog where he discusses technical topics, such as efficient matrix multiplication and the limits of backpropagation.
In his writing, Rahimi has called for a more systematic approach to benchmarking algorithmic fairness. In a 2021 essay for the OpenPanel newsletter, he argued that current evaluation metrics are insufficient for detecting bias in deployed systems. This essay was cited by regulators and researchers.
Selected Publications
Rahimi has authored over 60 peer-reviewed papers. Notable works include:
- Rahimi, A., and Recht, B. (2007). Random Features for Large-Scale Kernel Machines. NeurIPS.
- Forster, K., Rahimi, A., and Koller, D. (2010). "Data-Driven Calibration of a Sensor Network." Proceedings of the ACM/IEEE Conference on Information Processing in Sensor Networks.
- Rahimi, A., Recht, B., and Darrell, T. (2020). "Learning to Represent Programs with Properties." ICLR.
He has also co-edited a book, "Distributed Sensor Networks: A Multiagent Perspective," with his dissertation advisor, published by Springer in 2006.
Personal Life
Rahimi was born in Iran and immigrated to the United States with his family in the early 1980s. He holds dual citizenship and speaks Persian and English fluently. In interviews, he has cited his father, an engineer, as an early influence. He is an avid mountaineer and has participated in expeditions in the Himalayas, an activity he has discussed as informing his approach to risk assessment in engineering.