Education
- PhD in Computer Science
- MSc in Data Science
- BSc Applied Mathematics
- Image Processing and Optimization with MATLAB Summer School
Amir Mehrpanah
About
I am a PhD student in Computer Science at KTH Royal Institute of Technology, supervised by Hossein Azizpour, and co-supervised by Kevin Smith at SciLifeLab.
I am interested in theoretical foundations of explainability of deep networks.
I have been working on spectral analysis for understanding and explaining deep models trained on vision tasks.
I have completed my MSc in Data Science at Shahid Beheshti University of Tehran and my BSc in Applied Mathematics at Ferdowsi University of Mashhad.
This site collects news, publications, and maybe research notes.
Highlights
- Can we better explain classification with generative models?
Here I ask whether generative models can help explain decisions made by generative-based classifiers.
- Our latest paper on robustness of attribution methods is accepted at ICML 2026!
We show that SGD can improve robustness of gradient-based attribution, while attention-based attribution requires special care.
Resume Snapshot
Full CV
Academic Experience
- Teaching Assistant in Advanced Deep Learning
- Teaching Assistant in Probabilistic Graphical Models
- Teaching Assistant in Introduction to Machine Learning
- Teaching Assistant in Automata Theory
Selected Publications
All publications
Beyond Power Spectra: Cross-Frequency Interactions in Generative Dynamics
We demonstrate that cross-frequency interactions, rather than standard marginal spectral statistics, are essential for characterizing and distinguishing the learned dynamics of deep generative models.
Improving Adversarial Robustness of Attribution via Implicit Regularization
We analyze attribution robustness in CNNs and ViTs, showing that SGD can improve robustness of gradient-based attribution, while attention-based attribution requires kernelized attention for such gains.
On the Complexity-Faithfulness Trade-off of Gradient-Based Explanations
The paper proposes a spectral framework to analyze the smoothness–faithfulness trade-off in ReLU network explanations, formalizes the resulting explanation gap, and provides principled regularization with empirical validation.
Blog Posts
View blog
How Do You Use AI with X?
As an XAI researcher, I explain how I use AI with X.
The Era of AI-Assisted Coding and Its Hidden Technical Debt
I think we need to revise programming styles in the era of AI assisted coding.
Can we better explain classification with generative models?
Notes on using generative models to better explain classification decisions.
