The Machine INtelligence Development (MIND) Lab at the American University of Beirut (AUB) aims at providing state of the art solutions in the field of Artificial Intelligence with a focus on machine learning theory and applications. The lab has recently established a new identity, grew in size, and saw great progress in cutting-edge research in two main domains: Natural Language Processing and Context-Aware Sensing.

As we begin a new year, we are extremely proud of the diligence and perseverance that our lab has shown despite the challenges that oversaw the world and Lebanon the past year. Our students were able to push cutting-edge research, produce publications, compete in local and global challenges, and share their expertise with the world through open-source solutions and invited talks.

 

2020 Research Manifesto

 

Natural Language Processing (NLP)

Our NLP focused team has explored a wide range of challenges, from conversational systems and Arabic as a low-resource language to studying human reading comprehension. 

Gilbert Badaro, our lab’s newest Ph.D. graduate, has investigated multiple link prediction approaches and different natural language processing variations in order to try to link a large lexical Arabic Resource, SAMA, to English WordNet. In his work, he used a machine learning approach and developed training data using existing small-scale Arabic WordNet. His latest article is important because it provides publicly a large-scale sentiment lexical resource for Arabic, ArSenL 2.0, to enable more accurate sentiment mining models. It also presents a benchmark dataset to be used by other researchers in the field interested in automatically evaluating approaches for Arabic WordNet expansion. 

Obeida ElJundi has focused on exploring human reading comprehension. Reading comprehension is not the result of one single process as was thought before the 1970s. Rather, psychologists realized that a combination of several complex cognitive processes is involved. Traditional NLP algorithms used nowadays (e.g., RNN) cover some aspects of the reading comprehension cognitive processes, such as attention and forgetting, but render other aspects unaccounted for. Obeida has been looking into studying how the human brain reads and comprehends written text will help reveal what is missing and improve the overall performance of NLP.

 

Wissam Antoun has recently defended his thesis on Arabic Transformers for NLU and chatbots. This year, he trained the first Arabic BERT model, AraBERT, alongside his colleague, Fady Baly. Wissam is now planning to explore newer and better models for language understanding and generation, and he is off to a great start! Wissam and Fady have started 2021 by releasing 10 new Arabic Transformer Models (AraGPT2, AraELECTRA & AraBERTv2) on AUB Mind Lab and Hugging Face Github repositories.

 

 

Nataly Dalal has worked on designing lifelong learning for chatbots in customer support, thus allowing them to continuously learn and accumulate knowledge on their own from support forums posted on the web. Lifelong learning aims at producing chatbots that are more accurate, robust, and human-like. 

Christian Hokayem has focused the bulk of his research on conversational automated negotiation agents. With the goal of improving user experience in an automated sales setting, he investigated what empathy is and tried to find a way to incorporate it into agents’ language generation and decision-making models. In his work, he formulated the problem as a multi-objective reinforcement learning problem and taught the bot to maximize user experience (represented by sentiment) and sales price.

Tarek Naous, our newest master’s student, has started with research focused on enabling empathetic behavior in conversational agents for the Arabic Language with the global aim of making human-machine interactions more similar to human-human interactions. In 2020, he was able to release a dataset of empathetic conversations in Arabic and propose a neural-based model for generating empathetic responses. His work was published in WANLP 2020 and is currently in progress for further development, where the challenge is improving performance given the limited dataset size we have.

 

Context-Aware Sensing

With a focus on overcoming the challenge of learning from limited labeled data, our work in context-aware sensing has been versatile, ranging from applications of healthcare, such as traumatic brain injuries and epilepsy prediction, to power load forecasting and human activity recognition. 

Reem Mahmoud’s research interest is in personalized Machine Learning, which branches into problems of learning with little labeled data and advancing traditional transfer learning methods. This past year, she has looked into overcoming challenges of catastrophic forgetting in pre-trained neural networks on time-series applications, such as human activity recognition. Her work looked into improving performance on target tasks with limited training data while preserving performance on source tasks from the pre-trained model. This work, titled “Multi-Objective Learning to Overcome Catastrophic Forgetting in Time-series Applications,” has been recently submitted to the journal of ACM Transactions on Knowledge Discovery from Data. 

Marc Djandji has focused his research for the past year on improving the performance of deep learning models when trained with limited labeled training examples. He has specifically investigated the problem under the application of short-term load forecasting. His work looks into scalable multi-task learning methods.

This year, Alaa Khaddaj had completed an internship at Signal Kinetics (MIT Media Lab) under the supervision of Prof. Fadel Adib, where they developed an apparatus that can detect contaminated food or counterfeit products. The system used a generative model (variational autoencoder) to generalize to new unseen environments and a transfer learning scheme to learn from multiple experiments simultaneously. Alaa’s other line of research was to develop a domain adaptation model that can generalize a classifier between two similar domains that have some differences. For his final year project, his team developed an intelligent question-answering system for customer support. This work has been submitted as a journal paper.

Mosbah Aouad has worked on developing deep learning models for short-term load forecasting for residential houses, where the challenge is to deal with the high non-linearities present in the load data. To build a sensitive model to varying load patterns and large peaks in the data, he designed an attention mechanism to augment a sequence-to-sequence network to capture these variations. On the other side, his thesis work has focused on biomedical image analysis with the aim of predicting the survival rate of penetrating traumatic brain injury (pTBI) patients from brain CT scans analysis. He is designing a representation learning approach that captures relevant features reflecting the severity of pTBIs from CT scan analysis directly.

 

Open-source Solutions, Competitions & Other Activities

Beyond the efforts in research, we are proud of the efforts our team has put into sharing and pushing their expertise through participating in competitions, open-sourcing solutions, and delivering talks.

 

Here are some of our highlights from 2020:

 

2020 Graduates

    • Dalia Jaber – M.E. in Electrical & Computer Engineering
    • Wissam Antoun – M.E. in Electrical & Computer Engineering
    • Amir Hussein – M.E. in Electrical & Computer Engineering
    • Alaa Khaddaj – B.E. in Electrical & Computer Engineering
    • Gilbert Badaro – Ph.D. in Electrical & Computer Engineering
    • Raslan Kain – M.E. in Electrical & Computer Engineering

 

2020 List of Publications

Journal Papers

Conference Papers

 

To discover our lab and potential opportunities, visit our site or stay up to date with our news on LinkedIn and Twitter.