Amir H. Karimi
Assistant Professor
O'Donovan Chair in Trustworthy AI
Electrical & Computer Engineering, University of Waterloo
Cheriton School of Computer Science (cross-appointment)
Faculty Affiliate, Vector Institute
amirh.karimi [at] uwaterloo.ca
Bio
Dr. Amir-Hossein Karimi is an award-winning researcher and educator, and an Assistant Professor in the Department of Electrical and Computer Engineering and the Cheriton School of Computer Science (cross-appointed) at the University of Waterloo, and a Vector Institute Faculty Affiliate.
He leads the Collaborative Human-AI Reasoning Machines (✨CHARM✨) Lab, dedicated to advancing safe and trustworthy human-AI collaborations.
Having immigrated five times across three continents for studies and work, from Iran to Canada, and then onward to the USA, Germany, Switzerland, and the UK, Dr. Amir-Hossein Karimi has accumulated over 15 years of technical experience in both research, industry, and startup roles.
Prior to joining Waterloo, Dr. Karimi accumulated significant industry experience at leading tech companies such as
BlackBerry (2013),
Meta (Facebook) (2014-6),
Google Brain (2021),
and DeepMind (2022),
and provided consulting services for various startups and incubators including for NEXT AI.
His contributions have earned him multiple accolades, such as the UofToronto Spirit of Engineering Science Award (2015), the UWaterloo Alumni Gold Medal Award (2018), the NSERC Canada Graduate Scholarship - Doctorate (2018), the Google PhD Fellowship (2021), the ETH Zurich Medal (2024), the NSERC Discovery Grants & Supplements (2024), and the Igor Ivkovic Teaching Excellence Award (2024). To view the full CV, click here.
✨CHARM Lab✨ 👈🏼 click me
As AI becomes more embedded in everyday life, e.g., healthcare, education, finance, and transportation, our dependency on these systems grows, and so does the risk of potential consequential errors. The mandate of the Collaborative Human-AI Reasoning Machines (✨CHARM✨) Lab is to enhance the integration of AI systems into human decision-making, to build systems that allow users to understand 🔍, challenge ⚖️, and improve 🚀 AI decisions. The CHARM Lab brings expertise in explainable AI, causal inference, and neuro-symbolic approaches, and collaborates with leading experts in such fields as social psychology, cognitive science, human-computer interaction, multi-agent reinforcement learning, game theory, and behavioral economics.
Prospective Students
The lab is always on the lookout for exceptional and highly motivated students/visitors across all levels (bachelor's, master's, doctoral, postdoctoral). Our research sits at the intersection of theory and practice, focusing on foundational questions motivated by real-world impact—particularly in explainable AI, algorithmic recourse, neuro-symbolic systems, and machine metacognition. If this direction resonates with your interests and background in machine learning, computer science, or related fields, please fill out fill out this form.
Amir-Hossein Karimi
Principal Investigator (PI)
Mohammad Hadi Sepanj
Postdoctoral Fellow
Ahmed Abdelaal
PhD Student
Zahra Khotanlou
Master's Student
Hashir Ahmed
Master's Student
Chenghao Tan
Master's Student
Tom Wielemaker
Undergraduate Student
Hamza Mostafa
Undergraduate Student
Dongzhuyuan Lu
Master's Student
Jieming Yu
Master's Student
Abubakar Bello
Undergraduate Student
(next: Microsoft Inc.)
Mohammadreza Alavi
Undergraduate Student
Publications
Dr. Karimi's scholarly contributions have been showcased almost exclusively at top-tier AI and ML venues including NeurIPS, ICML, AAAI, AISTATS, ACM FAccT, and ACM AIES. He has authored influential publications such as a comprehensive survey paper in the prestigious ACM Computing Surveys, holds a patent, and is a contributing author of a book chapter. Dr. Karimi’s work on algorithmic recourse has notably elevated its prominence in responsible AI research, with its presence growing from almost none to hundreds on Google Scholar in just five years; algorithmic recourse is now a mandatory criterion in key sectors, including Canada’s Treasury Board Directive on Automated Decision-Making. Several of Dr. Karimi's papers have received over 100 citations each. Dedicated to knowledge mobilization and reproducibility, his open-source code has garnered over 100 GitHub stars.
Most recent publications are available on Google Scholar.
‡ indicates equal contribution.
🏛️ EurIPS Theory of XAI Workshop
✨ TL;DR: Solving XAI requires us to make explicit our causal assumptions of the world
🏛️ NeurIPS Workshop on Algorithmic Collective Action
✨ TL;DR: Providing recourse to any one individual affects what recourse can be provided to other individuals
🏛️ Trends in Cognitive Sciences (impact factor: 17.2)
✨ TL;DR: AI systems should integrate metacognitive strategies to enhance robustness, explainability, cooperation, and safety in complex real-world scenarios
🏛️ Nature Scientific Reports (impact factor: 3.9)
✨ TL;DR: TCFormer leverages CNNs, transformers, and temporal convolutions to decode EEG motor imagery, significantly outperforming existing BCI approaches
🏛️ ICML 2023 (acceptance rate: 27.9%)
✨ TL;DR: do explanations actually explain the predictions? yes, among other things....
🏛️ ACM Computing Surveys 2022 (impact factor: 24.0)
✨ TL;DR: definitions, formulations, solutions, and prospects for recourse research
🏛️ EMNLP 2016 (acceptance rate: 24.3%)
✨ TL;DR: use key-value pairs to efficiently retrieve and store information for improved document comprehension
🏛️ EurIPS Theory of XAI Workshop
✨ TL;DR: Solving XAI requires us to make explicit our causal assumptions of the world
🏛️ Trends in Cognitive Sciences (impact factor: 17.2)
✨ TL;DR: AI systems should integrate metacognitive strategies to enhance robustness, explainability, cooperation, and safety in complex real-world scenarios
🏛️ ICML 2024 (acceptance rate: 27.5%)
🏛️ ICML 2023 (acceptance rate: 27.9%)
✨ TL;DR: do explanations actually explain the predictions? yes, among other things....
🏛️ ACM AIES 2021 (acceptance rate: 38.0%)
✨ TL;DR: generate optimal counterfactual explanations using mixed-integer linear programs
🏛️ AISTATS 2019 (acceptance rate: 32.4%)
✨ TL;DR: generate optimal counterfactual explanations using satisfiability solvers
🏛️ NeurIPS Workshop on Algorithmic Collective Action
✨ TL;DR: Providing recourse to any one individual affects what recourse can be provided to other individuals
🏛️ ACM FAccT 2023 (acceptance rate: 24.6%)
✨ TL;DR: as the title says...
🏛️ ICML 2022 (acceptance rate: 21.9%)
✨ TL;DR: robustness of recourse is separate from robustness of prediction
🏛️ AAAI 2022 (acceptance rate: 15.0%)
✨ TL;DR: fairness of recourse is separate from fairness of prediction
🏛️ ACM Computing Surveys 2022 (impact factor: 24.0)
✨ TL;DR: definitions, formulations, solutions, and prospects for recourse research
🏛️ NeurIPS 2020 (acceptance rate: 20.1%)
✨ TL;DR: causal assumptions are rarely available; instead generate recourse probabilistically
🏛️ ACM FAccT 2021 (acceptance rate: 25.0%)
✨ TL;DR: real-world causal assumptions are needed for counterfactual explanations & recourse
🏛️ ACM AIES 2021 (acceptance rate: 38.0%)
✨ TL;DR: generate optimal counterfactual explanations using mixed-integer linear programs
🏛️ AISTATS 2019 (acceptance rate: 32.4%)
✨ TL;DR: generate optimal counterfactual explanations using satisfiability solvers
🏛️ EurIPS Theory of XAI Workshop
✨ TL;DR: Solving XAI requires us to make explicit our causal assumptions of the world
🏛️ AAAI 2023 (acceptance rate: 23.7%)
🏛️ ICML 2023 (acceptance rate: 27.9%)
✨ TL;DR: do explanations actually explain the predictions? yes, among other things....
🏛️ ICML 2023 (acceptance rate: 27.9%)
🏛️ ACM FAccT 2023 (acceptance rate: 24.6%)
✨ TL;DR: as the title says...
🏛️ ICML 2022 (acceptance rate: 21.9%)
✨ TL;DR: robustness of recourse is separate from robustness of prediction
🏛️ AAAI 2022 (acceptance rate: 15.0%)
✨ TL;DR: fairness of recourse is separate from fairness of prediction
🏛️ ACM Computing Surveys 2022 (impact factor: 24.0)
✨ TL;DR: definitions, formulations, solutions, and prospects for recourse research
🏛️ NeurIPS 2020 (acceptance rate: 20.1%)
✨ TL;DR: causal assumptions are rarely available; instead generate recourse probabilistically
🏛️ ACM FAccT 2021 (acceptance rate: 25.0%)
✨ TL;DR: real-world causal assumptions are needed for counterfactual explanations & recourse
🏛️ Trends in Cognitive Sciences (impact factor: 17.2)
✨ TL;DR: AI systems should integrate metacognitive strategies to enhance robustness, explainability, cooperation, and safety in complex real-world scenarios
🏛️ arXiv
🏛️ CVIS 2017 (acceptance rate: 40.0%)
🏛️ CCN 2017 (acceptance rate: 40.0%)
🏛️ Nature Scientific Reports (impact factor: 3.9)
✨ TL;DR: TCFormer leverages CNNs, transformers, and temporal convolutions to decode EEG motor imagery, significantly outperforming existing BCI approaches
🏛️ CVIS 2018 (acceptance rate: 40.0%)
🏛️ ICIAR 2017 (acceptance rate: 40.0%)
🏛️ Biomedical Optics Express (impact factor: 2.9)
🏛️ IJCNN 2018 (acceptance rate: 22.7%)
🏛️ EMNLP 2016 (acceptance rate: 24.3%)
✨ TL;DR: use key-value pairs to efficiently retrieve and store information for improved document comprehension
🏛️ ICIP 2016 (acceptance rate: 40.0%)
Teaching
Press, Funding, & Consultations
Dr. Karimi is grateful for the generous funding support from the University of Waterloo, NSERC, CIFAR, Google, Waterloo.AI, and the O'Donovan Family, enabling his team to push the boundaries of human-AI research.
Besides academic contributions, Dr. Karimi also consults internationally for several startups, leveraging his expertise for discovery, product-market fit, and scaling of operations, in addition to fundraising and grant writing. He is available to discuss how his research and experience can deliver value to your organization and stakeholders. To book a time, click this link.
For press inquiries, feel free to reach out via email.