Mohammad Mahdi Derakhshani

Computer Vision | Machine Learning

Hi! I'm Mohammad, a Ph.D. student of VIS lab at the University of Amsterdam working with Cees Snoek and Yuki Asano. My Ph.D. thesis is Efficient Adaptation of Large-scale Vision and Language Models. Specifically, my current focus is on the implementation of prompting and chain of thoughts for multi-modal reasoning.

In summer 2022, I was also a ML/CV intern at Samsung AI Center (SAIC) in Cambridge working with Brais Martinez and Georgios Tzimiropoulos doing research on Prompting of large-scale language-image models and Federated Learning.

Previously I was a master student at the University of Tehran working with Babak Nadjar Araabi and Mohammad Amin Sadeghi. I did my master at machine learning and computational modeling lab working on object detection and image compression. During my master degree, I also did research on object detection with Mohammad Rastegari at Allen AI Institute.

I am a member of ELLIS society, and I have served as reviewer for ICCV'21, ICML'22, CVPR'23, ICLR'23, and TPAMI.

News

TA for Deep Learning 2022 by Yuki Asano.
New preprint: variational prompt tuning improves generalization of vision and language models.
Joined Samsung AI Center (SAIC) as CV/ML summer intern in Cambridge in UK.
One paper accepted to MICCAI'22: introducing LifeLonger, a continual learning benchmark for disease classification.
TA for Deep Learning 2021 by Xiantong Zhen.
New preprint: introducing generative kernel continual learning.
One paper accepted to ICML'21: introducing kernel continual learning.
TA for Machine Learning 2020 by Eric Nalisnick.
Staring my PhD in January 2020
One paper accepted to CVPR'19: introducing Assisted Excitation of Activations for object detection.
One paper accepted to CVPR'18 (Workshop): introducing BlockCNN for artifact removal and image compression.

Research

I'm interested in computer vision and multi-modal learning (vision & language models), specifically the prompting and chain of thoughts for multi-modal reasoning.

Variational prompt tuning improves generalization of vision-language models

Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees G. M. Snoek, Georgios Tzimiropoulos, Brais Martinez.

arXiv 2022 bibtex

We propose a probabilistic modeling of the underlying distribution of prompts, allowing prompts within the support of an associated concept to be derived through stochastic sampling. This results in a more complete and richer transfer of the information captured by the language model, providing better generalization capabilities for downstream tasks.

LifeLonger: A Benchmark for Continual Disease Classification

Mohammad Mahdi Derakhshani, Ivona Najdenkoska, Tom van Sonsbeek, Xiantong Zhen, Dwarikanath Mahapatra, Marcel Worring, Cees G. M. Snoek.

MICCAI 2022 bibtex

We introduce LifeLonger, a benchmark for continual disease classification on the MedMNIST collection, by applying existing state-of-the-art continual learning methods. We perform a thorough analysis of the performance and examine how the well-known challenges of continual learning, such as the catastrophic forgetting exhibit themselves in this setting.

Generative Kernel Continual learning

Mohammad Mahdi Derakhshani, Xiantong Zhen, Ling Shao, Cees G. M. Snoek

arXiv 2021 bibtex

We introduce generative kernel continual learning, which explores and exploits the synergies between generative models and kernels for continual learning.

Kernel Continual learning

Mohammad Mahdi Derakhshani, Xiantong Zhen, Ling Shao, Cees G. M. Snoek

ICML 2021 bibtex

This paper introduces kernel continual learning, a simple but effective variant of continual learning that leverages the non-parametric nature of kernel methods to tackle catastrophic forgetting.

Assisted Excitation of Activations: A Learning Technique to Improve Object Detectors

Mohammad Mahdi Derakhshani, Saeed Masoudnia, Amir Hossein Shaker, Omid Mersa, Mohammad Amin Sadeghi, Mohammad Rastegari, Babak N. Araabi

CVPR 2019 bibtex

We present a simple and effective learning technique that significantly improves mAP of YOLO object detectors without compromising their speed.

BlockCNN: A Deep Network for Artifact Removal and Image Compression

Danial Maleki, Soheila Nadalian, Mohammad Mahdi Derakhshani, Mohammad Mahdi Derakhshani, Mohammad Amin Sadeghi

CVPR (Workshop) 2018 bibtex

We present a general technique that performs both artifact removal and image compression. For artifact removal, we input a JPEG image and try to remove its compression artifacts.