Pedram Khorsandi

Tight Lower Bounds and Improved Convergence in Performative Prediction

• Preprint (2024)

This paper extends the Repeated Risk Minimization (RRM) framework by leveraging historical datasets from previous retraining snapshots. It introduces a class of algorithms called Affine Risk Minimizers, proving the tightness of both new and existing bounds in the same regime, and demonstrating empirical improvements in convergence for performative prediction.

Are socially-aware trajectory prediction models really socially-aware?

• Elsevier Volume 141 (2022)

Investigates the adversarial robustness of trajectory prediction models by introducing a socially-attended attack that exposes their limitations in collision avoidance. The proposed method applies hard- and soft-attention mechanisms to guide perturbations, revealing gaps in social understanding while also enhancing model performance.

Do explanations explain the model? Model knows best

• CVPR (2022)

Introduces a framework for evaluating feature attribution methods by leveraging the neural network itself. The framework generates input features that induce specific behaviors in the output, enabling controlled experiments to assess whether an explanation method aligns with key axioms. The evaluation exposes properties and limitations of existing explanation techniques, providing insights into their reliability.

University de Montreal

Ph.D. in Computer Science • 2022 - Ongoing

Sharif University of Technology

Bachelor of Computer Science & Engineering • 2017 - 2022

Technical University of Munich (TUM)

Research Intern • 2021

Studied the trustworthiness of methods that explain neural networks’ decisions and proposed a benchmark to evaluate such methods in terms of accuracy, class sensitivity, and other aspects.

Visual Intelligence for Transportation (EPFL)

Research Intern • 2020

Introduced socially-attended attacks to shed light on the limitations of current trajectory prediction models and provided insights into making them more robust.

• I bike every weekend during the summer. I also share my art works and projects on my instagram (linked above). If you're interested in buying my artwork, feel free to email me ;).