Commonsense injection in large language models
|Hour||11:00 › 12:00|
|Speaker||Niket Tandon is a Senior Research Scientist at the Allen Institute for AI in Seattle. His research interests are in commonsense reasoning and natural language guided reasoning. He works at the Aristo team responsible for creating AI which aced science exams. He obtained his Ph.D. from the Max Planck Institute for Informatics in Germany in 2016, where he was supervised by Professor Gerhard Weikum, resulting in the largest automatically extracted commonsense knowledge base at the time, called WebChild. He is also the founder of PQRS research, an organization providing research opportunities to undergraduate students from underrepresented institutes. Homepage: https://niket.tandon.info/|
|Category||Conferences - Seminars|
Abstract: Large LMs, while powerful, are not immune to mistakes which are obvious to humans, but they can be prohibitively costly to retrain. Our goal is to effectively correct language model mistakes by injecting knowledge via user interactions with the system but without retraining. Our approach is a memory-augmented architecture, where user feedback is used to make the model generate a better answer or to post hoc correct the errors such that the model does not repeat similar mistakes. We will discuss efficient solutions to designing this memory of knowledge, and leveraging it in the model. This is a step in the direction of never ending learning, and we will present a future roadmap to what open research problems need to be addressed to get to never ending learning language models.