Welcome to My Homepage :-)

My name is Negar Foroutan (Persian: نگار فروتن), and I am from Iran. Currently I am a PhD student at the School of Computer and Communication Sciences (IC), EPFL, and a doctoral research assistant at LSIR and NLP labs under supervision of Karl Aberer and Antoine Bosselut.

I've got my Master degree in Artificial Intelligence at the Department of CSE and IT of Shiraz University, where I also received my B.Sc. degree in Software Engineering. During my master studies, I was working in the area of social network analysis, and in my thesis, I investigated the problem of inferring the structure and dynamics of networks using information diffusion.

After finishing my master studies, and for three months, I was a research intern at the Max Planck Institute for Software Systems (MPI-SWS) in Germany, under the supervision of Manuel Gomez Rodriguez. I've also had a chance to spend three months in Switzerland as an intern at the Data Analytics Laboratory, ETH Zurich, under the supervision of Carsten Eickoff.

I also spent a year working as a research assistant at the Machine Learning and Optimization laboratory (MLO) at EPFL, under the supervision of Martin Jaggi.

Research Interests: My research interests lie broadly in the fields of natural language processing (NLP) and machine learning. I am particularly interested in multilingual NLP and cross-lingual transfer, focusing on low-resource scenarios.

Education

Publications

  • C. Fierro, N. Foroutan, D. Elliott, and A. Søgaard, "How Do Multilingual Models Remember? Investigating Multilingual Factual Recall Mechanisms," arXiv, 2024. (More Information)

  • D. Bayazit, N. Foroutan, Z. Chen, G. Weiss, and A. Bosselut, "Discovering Knowledge-Critical Subnetworks in Pretrained Language Models," EMNLP, 2024. (More Information)

  • B. Borges*, N. Foroutan*, D. Bayazit*, A. Sotnikova*, et. al., "Could ChatGPT get an Engineering Degree? Evaluating Higher Education Vulnerability to AI Assistants," arXiv 2024 . (More Information)

  • N. Foroutan, M. Banaei, Karl Aberer, and A. Bosselut, "Breaking the Language Barrier: Improving Cross-Lingual Reasoning with Structured Self-Attention," EMNLP 2023 - Findings. (More Information)

  • Y. Karoui, R. Lebret, N. Foroutan, and K. Aberer, "Stop Pre-Training: Adapt Visual-Language Models to Unseen Languages," ACL, 2023. (More Information)

  • N. Foroutan, M. Banaei, R. Lebret, A. Bosselut, and K. Aberer, "Discovering Language-neutral Sub-networks in Multilingual Language Models," EMNLP, 2022. ( More Information)

  • N. Foroutan, A. Romanou, S. Massonnet, R. Lebret, and K. Aberer, "Multilingual Text Summarization on Financial Documents," Proceedings of the 4th Financial Narrative Processing Workshop@ LREC, 2022. ( More Information)

  • N. Foroutan and M. Jaggi, "Sparse Communication for Training Deep Networks," Workshop on Beyond first-order methods in ML systems at ICML, 2020. (More Information)

  • N. Foroutan and A. Hamzeh, "Discovering the Hidden Structure of a Social Network: A Semi Supervised Approach," IEEE Transactions on Computational Social Systems, pp. 14-25. 2017. ( More Information )

  • N. Foroutan Eghlidi, Jannick Griner, Nicolas Mesot, Leandro von Werra and Carsten Eickhoff, "ETH Zurich at TREC Precision Medicine 2017," Proceedings of the 26th Text Retrieval Conference (TREC), 2017.

  • N. Foroutan Eghlidi, A. Afshar, B. Ashenagar and A. Hamzeh, "A lightweight method to investigate unknown social network structure." Computer and Knowledge Engineering (ICCKE), 2015 5th International Conference on, pp. 262-267. IEEE, 2015. (More Information)

  • B. Ashenagar, N. Foroutan Eghlidi, A. Afshar and A. Hamzeh, "Team formation in social networks based on local distance metric." Fuzzy Systems and Knowledge Discovery (FSKD), 2015 12th International Conference on, pp. 946-952. IEEE, 2015. (More Information)

  • B. Ashenagar, A. Hamzeh, N. Foroutan Eghlidi and A. Afshar, "A fast approach for multi-objective team formation in social networks." Information and Knowledge Technology (IKT), 2015 7th Conference on, pp. 1-6. IEEE, 2015. (More Information)

  • A. Afshar, B. Ashenagar. N. Foroutan Eghlidi, Mansour Zolghadri Jahromi and Ali Hamzeh, "Using local utility maximization to detect social networks communities." Computer Science and Software Engineering (CSSE), 2015 International Symposium on, pp. 1-8. IEEE, 2015. ( More Information)

Projects

More information about my projects: LinkedIn

  • Discovering Language-neutral Sub-networks in Multilingual Language Models2022
    In this project, we investigate the effect of language-neutral parameters on the cross-lingual transfer performance of multilingual language models. We conceptualize language neutrality of multilingual models as a function of the overlap between language-encoding sub-networks of these models. We employ the lottery ticket hypothesis to discover sub-networks that are individually optimized for various languages and tasks. Our evaluation across three distinct tasks and eleven typologically-diverse languages demonstrates that sub-networks for different languages are topologically similar, making them effective initializations for cross-lingual transfer with limited performance degradation.

  • Cross-lingual Representation Learning for Low-Resource LanguagesMarch 2020 - June 2020
    In this project, our goal is to transfer a pre-trained monolingual model in a source language to a target language by zero-shot learning. Unlabelled data was used to train token embeddings for the target language, and labelled data from the source language was used to fine-tune the model for the down-stream task. In our experiments, we use the German language as the source and Swiss German as the target language. We evaluate the obtained model using sentiment analysis as our evaluation down-stream task.

  • News Article Contextual EmbeddingSeptember 2020 - December 2020
    In this project, we propose a method for news article embedding such that the new embedding space can categorize articles based on the events they cover. We use BERT embedding as a basis and add a feed-forward neural network on top of it. We also take advantage of triplet loss in the training process. Experimental results show the proposed method outperforms the baselines in terms of adjusted mutual information and adjusted rand index scores.

  • Art Creation Using Generative Adversarial Networks (GANs)October 2020 - December 2020
    In this project, we used BigGAN with inversion technique and progressive-growing GAN networks to generate images of Mesopotamian artifacts as well as Assyrian and (Neo-) Sumerian artifacts which evolved from the Mesopotamian region.

  • mlbench: Distributed Machine Learning BenchmarkSeptember 2018 - September 2019
    The goal of this project is to provide a collection of reference implementations of distributed machine learning algorithms for different frameworks and system platforms. Currently, we are working on supervised machine learning algorithms, such as deep learning tasks and linear models. We provide a set of defined tasks and related datasets to have a fair and precise comparison of all algorithms, frameworks, and hardware platforms.

  • TREC 2017 Precision Medicine trackAugust 2017
    The goal of this project was to retrieve biomedical articles and clinical trials addressing appropriate treatments for a given patient. We began by performing literal query term matching, taking into account the likelihood of document relevance in terms of cancer types, genetic variants, and demographics. Next, in a subsequent phase, we re-ranked the most relevant results based on a range of deep neural gene embeddings that project literal genetic expressions into a semantics-preserving vector space. We used feed-forward networks trained on PubMed and NCBI information but also relying on generative adversarial methods to determine the likelihood of co-occurrence of various mutations within the same patient/article. Experimental results show that even without existing expert labels, the proposed method can introduce marginal improvements over competitive TF-IDF baselines. I was involved in this project during my internship at ETH Zurich.

  • Citation Network AnalysisMarch 2017 - Present
    The goal of the project is to gain a better understanding of what happening in citation networks. Indeed, we want to quantify the value of a set of published papers and model the knowledge diffusionin a citation network. First, we assume that writing a paper is an event which is generated by acounting process. Then, it is supposed that there is an infinite number of latent of cluster each event belongs and each cluster has a characteristic vocabulary and cited authors. Given a citation network contains a set of papers, and by using the words in the papers and the authors they cite, we want to infer the latent clusters. These clusters will reveal something meaningful about research trends, cited authors, citation process, etc. This project is a collaboration work that I was involved during my stay at the MPI-SWS.

  • Inferring Social Networks Structure (M.Sc. Thesis) 2014-2016
    Diffusion of information, spread of rumors, ideas and diseases are assumed to be stochastic processes that occur over the edges of some network structure. Many times, the underlying network structures are unobserved and we only observe the infection times of nodes. In some cases, such networks are also dynamic and change over time. In my thesis, I investigated the problem of inferring the structure and dynamics of networks using information diffusion. First, I modeled the diffusion network as a Markov Decision Process (MDP). Then, as Reinforcement Learning (RL) is one the methods to solve finite MDPs, I used Q-learning, which is a kind of RL, to solve this MDP.

  • Face Identification and Tracking from a Video Source December 2016
    The goal of this project is to identify and to track faces in a live video stream. Here, we have to train a model for each face we want to identify. To do this, we use different feature extraction and selection methods to find features which best explain the faces. Also, some normalization techniques are applied to normalize a face for position and illumination to reduce the variance caused by these.

  • Predicting Excitement for DonorsChoose.org KDDCup2014 July 2014
    The task of KDD Cup 2014 was predicting the excitement of available projects on DonorsChoose.org. We employed some data mining approaches in order to extract relevant and discriminative features and we studied some sampling methods to solve imbalanced data problem. We predicted project's excitement using some models based on Random Forest, Stacking, and Naive Bayes. Also, we were at Top 25% of participants in this competition.

  • Perception of Emotion in Facial Expressions February 2014
    In this project, we want to design a system to classify facial expressions of emotion. After employing some image processing techniques to extract initial features we used epsilon-SVR and kernel methods to map features into another space to find more discriminative and shape-related features. We model emotion detection using an SVM classifier that calculated the probability for each emotion category.

  • Information Retrieval System July 2014
    The goal of the project is to Implement a vector-space information retrieval system. By knowing the set of relevant and non-relevant documents for a given query we tried to find optimal separator and improve it's performance by maximizing F-measure.

  • Malware Detection February 2014
    In this project we implemented a system to distinguish malicious from benign binary files. Extracting the n-gram of API Call sequences as our dataset, we employed two different methods to solve this problem: 1. "XCS" which is a reinforcement learning-based Genetic algorithm, and 2. N-gram based feature weighting using Genetic Algorithm.

  • Fuzzy Rule-Based Classification System February 2014
    The goal of this project is to learn rule weights in fuzzy rule-based classification systems. After constructing an initial rule-base for the problem, we applied an algorithm based on hill-climbing search method in which sequentially the solution improved by finding a neighbor solution that is better than the current one. This approach is fast, efficient, and tries to keep much fewer rules than the initial one.

Contact Me

The best way to contact me is via email, but you can find me on LinkedIn as well.

Email    Email: negar [dot] foroutan [at] epfl [dot] ch
LinkedIn    LinkedIn: negarforoutan
GitHub   GitHub: negar-foroutan
Twitter   Twitter: negarforoutan
Google Scholar   Google Scholar