top of page
me.HEIC

Ashia C. Wilson

I am a Lister Brothers Career Development Assistant Professor at MIT. My research focuses on designing scalable, reliable and socially responsible AI systems using tools from dynamical systems theory, statistics, and optimizationHere is a list of some information about me (CV, Publications, Contact) Short Bio: I obtained my B.A. from Harvard with a concentration in applied mathematics and a minor in philosophy, and my Ph.D.  from UC Berkeley in statistics. Before joining MIT, I held a postdoctoral position in the machine learning and statistics group at Microsoft Research.  ​​

Affiliations 

Teaching

  • 6.401 - Introduction to Data Science and Statistics (Spring 2021)

  • 6.036 - Introduction to Machine Learning (Spring 2022, Spring 2023)

  • 6.3950 - AI, Decision Making and Society (Fall 2022, Fall 2023, Fall 2024 (syllabus))

 

Recent Publications & Working Papers

​​Efficiency of Optimization

  • Adaptive Backtracking for Faster Optimization

        Joint with Laurent Lessard and Joao Cavalcanti. 2024.

        [arXiv]

  • Fast sampling from constrained spaces using the Metropolis-adjusted Mirror Langevin Algorithm

        Joint with Vishwak Srinivasan and Andre Wibisono

        Conference on Learning Theory (COLT).  2024.

        [arXiv]

  • Mean-field underdamped Langevin dynamics and its spacetime discretization

        Joint with Qiang Fu 

        International Conference on Machine Learning (ICML). 2024.

        [arXiv]

  • Accelerated Stochastic Optimization Methods under Quasar-convexity 

        Joint with Qiang Fu and Dongchu Xu​

        International Conference on Machine Learning (ICML). 2023.

        [arXiv, code

  • Sufficient conditions for non-asymptotic convergence of Riemannian optimisation methods

        Joint with Vishwak Srinivasan

        OPTML Workshop, Neurips (Spotlight). 2022.

        [arXiv

        

​​Integrity of Predictions

  • Unstable Unlearning: The Hidden Risk of Concept Resurgence in Diffusion Models

        Joint with Vinith Suriyakumar, Rohan Alur, Ayush Sekhari and Manish Raghavan. 2024.

        [arXiv]

  • Faster Machine Unlearning via Natural Gradient Descent

        Joint with Omri Lev. 2024.

        [arXiv]

  • The Revealed Preferences of Pre-authorized Licenses and Their Ethical Implications for Generative Models

        Joint with Vinith Suriyakumar, Peter Menell and Dylan Hadfield-Menell

        The 2nd Workshop on Generative AI and Law (GenLaw). 2024.

        [link

  • Algorithms that Approximate Data Removal: New Results and Limitations   

        Joint with Vinith Suriyakumar

        Advances in Neural Information Processing Systems (NeurIPS). 2022.

        [arXiv, code] 

​​Evaluation of Decisions

  • Algorithmic Pluralism: A Structural Approach To Equal Opportunity

        Joint with Shomik Jain, Vinith Suriyakumar, and Kathleen Creel

       ACM Conference on Fairness, Accountability and Transparency  (FAccT, best paper). 2024.

       ACM conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO, non-archival). 2023. 

        [arXiv] 

  • Homogenous Algorithms Can Reduce Competition in Personalized Pricing

        Joint with Nathan Jo, Kathleen Creel, and Manish Raghavan

        ICML Workshop on Humans, Algorithmic Decision-Making and Society: Modeling Interactions and Impact. 2024.

  • Scarce Resource Allocations That Rely On Machine Learning Should Be Randomized

        Joint with Shomik Jain and Kathleen Creel​ 

        International Conference on Machine Learning (ICML). 2024.

        [arXiv]

  • As an AI Language Model, "Yes I Would Recommend Calling the Police": Norm Inconsistency in LLM Decision-Making

        Joint with Shomik Jain and Dana Calacci 

        Conference on AI, Ethics and Society (AIES). 2024. 

        [arXiv]

  • Automating Transparency Mechanisms in the Judicial System Using LLMs: Challenges and Opportunities 

        Joint with Ishana Shastri, Shomik  Jain and Barbara Engelhardt

       Conference on AI, Ethics and Society (AIES). 2024. 

       [arXiv

bottom of page