Talks & Seminars
Title: Neural-Symbolic Reasoning for Complex Question Answering Interactive AI
Ms. Amrita Saha, IBM Research
Date & Time: July 16, 2019 15:00
Venue: Department of Computer Science and Engineering, Room No. 109, 01st Floor, New CSE/CC Building
We will open this talk discussing about various complex setting of question answering, for e.g. visual conversation or complex conversational QA over KBs or learning to answer questions by reading a paraphrase comprehension. Then we will delve deeper into complex KBQA problem where the recent proliferation of deep learning has lead to end-to-end learning of neural representations of the query and its transformation to reach the answer. However such monolithic models can neither handle complex multi-step reasoning well nor provide any interpretable output at the intermediate stages to explain model's inferencing technique. Rather what we propose is a modular neural network which learns to decompose a complex question answering task into a sequence of simpler steps, called a "program". This paradigm called "neural program induction" has been applied in several settings in the recent past, but under certain constraints or assumptions; for e.g. either for a simpler question answering task or assumption that the "oracle program" is provided as part of supervision during training. Though this keeps the learning problem simple, these assumptions in practice can be very expensive and often, infeasible. Targeting these limitations, we introduce Complex Imperative Program Induction from Terminal Rewards (CIPITR) to answer complex questions (involving sequence of logical, quantitative and comparative reasoning) over a large scale KB Wikidata, with only final answer as distance supervision. The main challenge arises from the combinatorial search space of programs that arise from the absence of direct supervision, compounded by an extremely sparse reward space. We will talk about how CIPITR learns to pragmatically search only for "meaningful" programs in this exponential space by incorporating both generic and task-specific programming rules as symbolic constraints in its search. Marrying the symbolic AI with the neural reasoning in this way, CIPITR can exploit the best of both worlds and consequently succeeds in beating the competing QA models by a significant margin. We will conclude the talk discussing about the potential of CIPITR in pushing the frontiers of complex reasoning tasks in various other settings, like answering questions or conversing strategically over multiple modalities (like text/images/videos etc) in a more human-like fashion.
Speaker Profile:
Amrita Saha is a Research Engineer at IBM Research, where she has worked for over six years on various research problems at the intersection of language and vision, developing machine learning and deep representation learning techniques for learning to answer questions or converse or even debate over a mix of structured or unstructured multi-modal data. Her experience at IBM started with working on a futuristic Grand Challenge research on building an artificial Debater, to an industry-impacting research on bringing cognitive computing technologies to the world of fashion and more recently, to leading a broader research agenda on building interactive multimodal AI capabilities in collaboration with academia. She obtained her Masters degree in Computer Science from Indian Institute of Technology Bombay, India in 2012, prior to joining IBM Research. Over the years she has published in various reputed conferences and journals like TACL, ACL, AAAI, SDM, NeurIPS, COLING etc, as well as organized workshops in conferences like KDD, ICCV, and also filed multiple patents in various areas of AI.
List of Talks


Faculty CSE IT
Forgot Password
    [+] Sitemap     Feedback