Explicit Knowledge and Reasoning in the Era of Deep Learning

Seminar talk titled "Explicit Knowledge and Reasoning in the Era of Deep Learning"

Title Of the Talk: Explicit Knowledge and Reasoning in the Era of Deep Learning
Speaker: Dr Somak Aditya
Host Faculty: Dr. Vineeth N Balasubramanian
Date & Time: Wednesday, 2nd June 12:00pm
Seminar link: https://iith.webex.com/iith/j.php?MTID=m4a44bc1889287324589fe1ea447f422f


Lifelong acquisition and reasoning with knowledge is known to play an integral part in the way humans understand the universe. In a recent talk, historian Yuval Noah Harari attributes the rise of humans to the discovery of language and provides examples of how we use language to communicate knowledge that led to our evolution. I (among many others) believe that even in the era of learning everything from scratch, knowledge plays a fundamental role in machine understanding. In the first part of the talk, I will summarize my contributions during my Ph.D. Thesis related to knowledge integration towards image understanding tasks, such as visual question answering.
While the efforts towards knowledge integration is now becoming mainstream, one of the problems plaguing the deep pre-trained language models (mainly Transformers-based) is that they often fail at non-adversarial simpler examples, showing a lack of reasoning ability and logical consistency. During my post-doctoral tenure, I have worked with the hypothesis that different types of reasoning may require different models (may be symbolic); and our recent work in various tasks show the indication of moving beyond the one-model-to-reason-them-all strategy currently adapted by the community. In my second part of the talk, I will focus on such evaluation strategies and how they can influence model building. Finally, I argue that while we develop individual capabilities such as language and visual understanding, it's important to be cognizant that such models will be deployed amongst humans. And a key to (extrinsic) human-AI trust establishment is for humans to be able to anticipate AI model's behavior. I argue that overall knowledge and reasoning will play a key role in creating such agents.

Speaker Profile:

Dr. Somak Aditya (Microsoft Research India, Personal Webpage ) is currently a Postdoctoral Researcher at Microsoft Research India. He was awarded his Ph.D. for the dissertation “Knowledge and Reasoning for Image Understanding'' from Arizona State University in 2018. He has a background in NLP, Knowledge Representation and Reasoning, and Machine Learning; and is experienced in conducting interdisciplinary research in the field of AI. After demonstrating the need for knowledge and reasoning in vision and language applications, he has explored the same in the Marketing domain during his 1.5 years stint at Adobe Research, and then broadly in Natural Language Processing at Microsoft Research. Through workshops, Somak Aditya has actively promoted knowledge integration in the community. He has been a co-organizer for 3 workshops: the R2K workshop in KR 2018 in Phoenix, Arizona, the NSNLI workshop in IJCAI 2021, and the KINN workshop in CIKM 2021.

Wednesday, 2nd June 12:00pm