Farnam Mansouri

Computer Science · University of Waterloo

About

I am a Ph.D. student in Computer Science at the University of Waterloo, advised by Shai Ben-David and Sandra Zilles. I am also currently a part-time machine learning researcher at RBC Borealis.

I study the theory of machine learning, with a focus on questions inspired by real-world applications. In particular, I am interested in how having more or less supervision affects what can be learned.

Before joining Waterloo, I completed my master’s degree at the University of Toronto under the supervision of Murat Erdogdu and Amir-massoud Farahmand, working on theoretical reinforcement learning. I also interned at the Max Planck Institute for Software Systems, where I worked on machine teaching with Adish Singla.

Farnam Mansouri

Publications

Main research themes

Learning from positive (and unlabeled) data

What can we learn when negative examples are unavailable? I study learning from positive examples alone, or from positive examples together with unlabeled data.

Related papers: [P1] [P4] [P6] [M2]

Learning with multiple target labelers

I study learning settings in which correct labels can come from multiple sources, rather than a single target labeler. These settings are motivated by tasks such as question answering with large language models, where a question may have several valid answers.

Related papers: [P3] [M1] (and ongoing work on the sample complexity of learning from multiple hidden thinkers).

Machine teaching

In machine teaching, the training data—or additional examples supplied to the learner—are curated by a knowledgeable teacher. I study how this guidance affects learning.

Related papers: [P7] [P8] [P11] [P12] [M3] [M5] [M6]

All publications

Refereed publicationsPreprints and manuscripts

Refereed publications

  1. Surprises in Proper Positive-Only Learning

    S. Ben-David, F. Mansouri, A. Mehrotra, M. Zampetakis†

    The 40th Annual Conference on Neural Information Processing Systems (NeurIPS), 2026. Oral Presentation, Top 0.4% of Submissions.

  2. Near-Optimal Sample Complexity of Deterministic Agnostic PAC Learning

    S. Ben-David, S. Hanneke, F. Mansouri, A. Shaeiri†

    The 40th Annual Conference on Neural Information Processing Systems (NeurIPS), 2026.

  3. Learning with Multiple Correct Answers – Regret Bounds under Different Feedback Models

    A. F. Pour*, F. Mansouri*, S. Ben-David

    The 40th Annual Conference on Neural Information Processing Systems (NeurIPS), 2026.

  4. Active learning from positive and unlabeled examples

    F. Mansouri, S. Zilles, S. Ben-David

    The 40th Annual Conference on Neural Information Processing Systems (NeurIPS), 2026.

  5. The Computational Complexity of Almost Stable Clustering with Penalties

    F. Mansouri, S. Zilles, K. Khodamoradi

    The 35th International Joint Conference on Artificial Intelligence (IJCAI), 2026.

  6. Learning from positive and unlabeled examples—Finite size sample bounds

    F. Mansouri, S. Ben-David

    The 39th Annual Conference on Neural Information Processing Systems (NeurIPS), 2025.

  7. Formal Models of Active Learning from Contrastive Examples

    F. Mansouri, H. U. Simon, Y. Chen, A. Singla, S. Zilles

    The 39th Annual Conference on Neural Information Processing Systems (NeurIPS), 2025.

  8. On Batch Teaching with Sample Complexity Bounded by VCD

    F. Mansouri, H. U. Simon, A. Singla, S. Zilles

    The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. Spotlight Paper.

  9. Deep Reinforcement Learning for Online Control of Stochastic Partial Differential Equations

    E. Pirmorad*, F. Khoshbakhtian*, F. Mansouri*, A. M. Farahmand

    NeurIPS Workshop on the Symbiosis of Deep Learning and Differential Equations (DLDE), 2021. Spotlight paper.

  10. ChOracle: A Unified Statistical Framework for Churn Prediction

    A. Khodadadi, A. Hosseini, E. Pajouheshgar, F. Mansouri, H. R. Rabiee

    IEEE Transactions on Knowledge and Data Engineering (IEEE TKDE), 2020.

  11. Understanding the Power and Limitations of Teaching with Imperfect Knowledge

    R. Devidze*, F. Mansouri*, L. Haug, Y. Chen, A. Singla

    The 29th International Joint Conference on Artificial Intelligence (IJCAI), 2020.

  12. Preference-Based Batch and Sequential Teaching: Towards a Unified View of Models

    F. Mansouri, Y. Chen, A. Vartanian, X. Zhu, A. Singla

    The 33rd Conference on Neural Information Processing Systems (NeurIPS), 2019.

Preprints and manuscripts

  1. PAC Learning with Multiple Correct Answers

    S. Hanneke, F. Mansouri, A. F. Pour, A. Shaeiri, K. Wang†

    Submitted to the 59th Annual ACM Symposium on Theory of Computing (STOC), 2027.

  2. Deterministic Proper Positive-Only Learning is Independent of ZFC

    S. Ben-David, F. Mansouri, A. Mehrotra†

    Submitted to the 38th International Conference on Algorithmic Learning Theory (ALT), 2027.

  3. Query Feedback and Teaching for Stable and Certifiable Language Generation

    F. Mansouri, A. F. Pour, G. Velegkas, S. Zilles†

    Submitted to the 38th International Conference on Algorithmic Learning Theory (ALT), 2027.

  4. On the Convergence and Approximation of Policy Gradient Methods for Spectral Measures

    F. Mansouri

    M.Sc. Thesis, 2022.

  5. Learning Half-Spaces from Perturbed Contrastive Examples

    A. A. R. Ravari*, F. Mansouri*, Y. Chen, V. Ivarson, A. Singla, S. Zilles

    arXiv:2602.02080, 2026.

  6. Preference-Based Batch and Sequential Teaching

    F. Mansouri, Y. Chen, A. Vartanian, X. Zhu, A. Singla

    arXiv:2010.10012, 2020.

* Equal contributions. † Authors are listed in alphabetical order.

Awards

  • David R. Cheriton Graduate Scholarship, 2026–2028. 20,000 CAD
  • NSERC Canada Graduate Scholarship-Doctoral (CGS-D), 2024–2027. 120,000 CAD
  • David R. Cheriton Graduate Scholarship, 2023–2025. 20,000 CAD
  • Ontario Graduate Scholarship, 2020–2021. 15,000 CAD
  • Vector Institute Scholarship in Artificial Intelligence, 2020–2021. 17,500 CAD
  • Iran National Mathematics Olympiad, 2014. Silver medal; among the top 40 students in Iran, out of more than 30,000 initial participants.