Series to Highlight Young Professionals in Signal Processing: Dr. Hamid Palangi

This issue brings to you our interview with Dr. Hamid Palangi, a Senior Researcher at Microsoft Research Lab (AI) in Redmond, Washington. His current research interests are mainly in the areas of Natural Language Processing and Reasoning across Language and Vision.

He received his Ph.D. from the University of British Columbia in Canada where he mainly worked on Sparse Decomposition and Compressive Sensing. Dr. Palangi’s work that was published in 2015 on sentence embedding for web search engines and IR received the prestigious IEEE Signal Processing Society Best Paper Award (Test of Time) last year.

During the last 3 years, he has been working on leveraging Neuro-Symbolic Representations for Textual Question Answering [AAAI 2018 and NeurIPS 2017 Explainable AI Workshop], Reasoning Grounded in Images [ICML 2020 and CVPR 2020 VQA and Dialog Workshop], Solving Algebraic Math Problems [ICML 2020 and Best Paper Award at NeurIPS 2019 KR2ML Workshop], Image Captioning and Multimodal Information Retrieval [in close collaboration with Microsoft Bing] and Better Transfer among Various NLP Tasks [HUBERT].

Dr. Palangi has served as Area Chair for Multimodality Track at ACL 2020, is a member of the organizing committee for ACL 2020, a PC member of AAAI, reviewer of NeurIPS, ICML, CVPR, and several IEEE Transactions. He also works as a mentor at Microsoft AI School advanced projects class (AI-611, currently only available for Microsoft FTEs) and Microsoft AI Residency Program.

We approached Dr. Palangi with a few questions:

I was born in a small city in west of Iran called Hamadan. I was the youngest of 5 children, my father was an engineer mainly into radios and TVs (this is the time when they were built using lamps and vacuum tubes, no modern transistors or integrated circuits, at least not in our city) and my mother was a full-time mom.

I got introduced to modern electronics by my 3 brothers all of whom decided to pursue their career in electrical engineering, and, well, that was how I got into it. The only family member who decided not to be an engineer was my sister who loved psychology much more. Well, after deciding to be an engineer, I needed to work hard to enter a good undergraduate program, in Iran, you should participate in a national 4-5 hour exam which is very competitive, and that’s it, either you are on the top and you get to choose which school you go, or you are not and your options will be limited.

I was not on the top so my best option to be an electrical engineer and move to the capital (Tehran) where my brother was living was to go to a university that had a program with special requirement, you need to teach electronics to high school students after you graduate for a specific amount of time, it was called teacher-training university. It ended up to be a quite helpful program for me, on the top of electrical engineering courses we were required to pass several extra courses related to teaching, psychology, and education, from which I learned a lot. That encouraged me to become a double major and get a bachelor’s degree in something called TEFL (Teaching English as a Foreign Language) which both had more courses related to psychology/education and could help me to improve my English. For the M.Sc. national entrance exam, I got on the top of the list and was able to get into the best school in the country, Sharif. I was fortunate to be in great labs at Sharif with a positive and energetic environment.

I worked at a start-up for about 1 year after my M.Sc. as an engineer and moved to beautiful Vancouver in Canada to pursue my Ph.D. at UBC afterward. Three months into the program in 2012, there was a workshop in our department that I attended where Li Deng from Microsoft Research was presenting their recent results using this new kind of deep learning models for speech recognition with impressive results, that was how I got introduced to deep learning. Thanks to my Ph.D. advisors who gave me a lot of freedom, I took a journey during my Ph.D. where I worked on a diverse set of problems related to deep learning models, from compressive sensing to speech recognition, to web search and information retrieval. From the internships during my Ph.D. and after joining Microsoft Research as full-time employee, I got the opportunity to collaborate and learn from so many great talented colleagues at Microsoft and great researchers from different universities for which I am grateful.

Q. How does your work affect society?

With all the recent advances in artificial intelligence, it is very important to focus on building systems and models that can help and empower people to perform their daily tasks. For example, if you are searching for something in the pool of documents you have, trying to find the answer for a specific question without the need to read hundreds of documents, or finding the answers to questions that are grounded in a specific context, e.g., questions about what we see or saw at some point (questions grounded in images/videos in this case). We also need to make content accessible for people in need, for example, making visual content (image, video, presentation slides with images) accessible for the audience with visual impairment, e.g., by generating text descriptions that are grounded in the given visual content.

Beyond these capabilities, it is also important for model developers to be able to have an understanding of internal mechanisms inside the current huge black-box models to be able to both explain the process/reasoning for each decision they make, and trying to go back and fix them whenever necessary if possible. My current efforts on building and leveraging neuro-symbolic representations to perform these tasks are the early steps towards these directions that can help people eventually to achieve more.

Q. How your work is related to signal processing?

Signal Processing is the foundation of how we do the modeling and understand the physical world around us. From the first steps where we were taught what is Fourier transform and why it can give us a better understanding of a diverse set of phenomena by studying their frequency behaviors, to when various transformations including Wavelets and Curvelets were introduced, to the time when we were introduced to dictionary learning (e.g., K-SVD) and sparse decomposition world where we learned it is better to learn the atoms instead of only using a predefined basis function, to the modern era of neural networks and deep learning, we are leveraging all the modeling strategies we learned through our education in Signal Processing.  

Q. How has your involvement with the IEEE SPS helped you in building your career?

I remember the first ICASSP that I both attended and helped as a volunteer was in 2013 in Vancouver, Geoffrey Hinton gave a keynote there and said they are going to show that Hidden Markov Models (HMMs) can be totally replaced by Recurrent Neural Networks (RNNs), and they actually did that later on, impressive!! At the end of his talk, he showed results of training a character level RNN language model (a model that given previous characters can predict the next one) trained by his student Ilya Sutskever on Wikipedia data, by conditioning it on “The meaning of life is”, the model generated “the tradition of the ancient human reproduction: it is less favorable to the good boy for when to remove her bigger.” That was very inspiring for me and I think for many other graduate students and researchers attending the conference. Since then I have been attending various IEEE conferences /workshops either presenting a paper, giving a tutorial (e.g., my tutorial in IEEE GlobalSIP 2017), or giving a talk.

I have also been involved with several IEEE Transactions. All these have helped me to better understand and learn about future promising directions, find meaningful connections with other researchers during the conference, and initiating collaborations between academic research labs and our lab in the industry.

Q. Anything else that you would like to add? 

If you give a try to the more modern Transformer based neural language models like GPT-2 by conditioning it on “The meaning of life is” you get something like “not to have a perfect understanding of the world in its entirety, but to be able to perceive it for what it is”, significantly improved since 2013 …