Imad Eddine MAROUF

I am a PhD student at the Institut Polytechnique de Paris, specializing in the continual learning of deep learning models and parameter-efficient training methods. Under the supervision of Dr. Stéphane Lathuilière and Dr. Enzo Tartaglione at the LTCI Laboratory, Telecom Paris, my research delves into the development of efficient training approaches for large models and the application of continual learning principles to vision-language models.

Previously, I interned at the Computer Vision Center in Barcelona, working on continual learning for vision-language models under the supervision of Pr. Joost Van de Weijer. I hold a Master’s degree in Data & AI from Institut Polytechnique de Paris and an undergraduate degree in Computer Science and Electronics from Sorbonne Université.

I am actively seeking research internship opportunities in vision-language models, continual learning, or generative models for the year 2025. Feel free to contact me for collaboration or any research opportunities.

news

Oct 04, 2024	CoFiMA was presented as an Oral paper (Top 3%) at ECCV’24, Milan, Italy.
May 01, 2024	Started a research internship at Computer Vision Center (CVC) Lab in Barcelona, Spain, under the supervision of Joost van de Weijer.
Jan 03, 2023	Thrilled to share that my work MiMi will be presented at WACV 2024.

latest posts

Oct 18, 2024
Oct 18, 2024
Oct 18, 2024

Publications

Weighted Ensemble Models Are Strong Continual Learners

Imad Eddine Marouf, Subhankar Roy, Enzo Tartaglione, and 1 more author

In European Conference on Computer Vision (ECCV) , 2024

Abs HTML PDF Code

Continual learning (CL) aims to learn from a stream of tasks without forgetting previously acquired knowledge. While most CL methods focus on mitigating catastrophic forgetting, we argue that the key to strong CL performance lies in effectively leveraging knowledge from all tasks. In this work, we propose CoFiMA, a simple yet effective approach that combines fine-tuning and model averaging. CoFiMA maintains a separate model for each task and uses a weighted ensemble for inference. We introduce a novel weighting scheme based on Fisher Information, which effectively balances the contributions of different task models. Our extensive experiments on various CL benchmarks demonstrate that CoFiMA consistently outperforms state-of-the-art methods, often by a significant margin. We provide theoretical insights into why CoFiMA works well and empirically validate its effectiveness in different CL scenarios.
Mini but Mighty: Finetuning ViTs With Mini Adapters

Imad Eddine Marouf, Enzo Tartaglione, and Stéphane Lathuilière

In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , Jan 2024

Abs HTML PDF Code

Vision Transformers (ViTs) have shown remarkable performance across various computer vision tasks. However, finetuning these large models on downstream tasks can be computationally expensive and prone to overfitting. In this paper, we introduce Mini Adapters (MiMi), a lightweight and efficient approach for adapting ViTs to new tasks. MiMi inserts small, task-specific adapters into the ViT architecture, allowing for efficient finetuning while maintaining the pre-trained model’s knowledge. Our method significantly reduces the number of trainable parameters and computational cost compared to full finetuning, while achieving comparable or better performance. We demonstrate the effectiveness of MiMi on various image classification tasks and provide insights into its workings through extensive ablation studies.
Rethinking Class-incremental Learning in the Era of Large Pre-trained Models via Test-Time Adaptation

Imad Eddine Marouf, Subhankar Roy, Enzo Tartaglione, and 1 more author

Jan 2024

Abs HTML PDF Code

Class-incremental learning (CIL) aims to continually learn new classes while retaining knowledge of previously learned ones. Traditional CIL methods often struggle with catastrophic forgetting and are computationally expensive. In this paper, we propose a novel approach that leverages large pre-trained models and test-time adaptation for CIL. Our method, TTA-CIL, uses a frozen pre-trained backbone and only updates a small set of parameters during test time to adapt to new classes. This approach significantly reduces computational costs and memory requirements while maintaining high performance across all classes. We demonstrate the effectiveness of TTA-CIL on various benchmark datasets and compare it with state-of-the-art CIL methods.