Dr. Sameer Singh is an Associate Professor of Computer Science at the University of California, Irvine (UCI). He is working primarily on the robustness and interpretability of machine learning algorithms and models that reason with text and structure for natural language processing. Sameer was a postdoctoral researcher at the University of Washington and received his Ph.D. from the University of Massachusetts, Amherst. He has received the NSF CAREER award, UCI Distinguished Early Career Faculty award, the Hellman Faculty Fellowship, and was selected as a DARPA Riser. His group has received funding from Allen Institute for AI, Amazon, NSF, DARPA, Adobe Research, Hasso Plattner Institute, NEC, Base 11, and FICO. Sameer has published extensively at machine learning and natural language processing venues and received conference paper awards at KDD 2016, ACL 2018, EMNLP 2019, AKBC 2020, ACL 2020, and NAACL 2022. (https://sameersingh.org/)
Title : Editing Machines: Updating Text with Up-to-date Information
Abstract : Reflecting real-world facts and consistent, up-to-date information in text is an important problem. Human-written text may deviate from truth either intentionally, to mislead, or because it becomes obsolete as the world changes. This issue extends to language model-generated text as well, where inconsistencies may arise due to hallucinations or models becoming stale, reflecting an outdated world view.
In this talk, I will discuss our work, beginning with the introduction of a novel generation task called faithfully reflecting updated information in text (FRUIT). We will explore the FRUIT-WIKI dataset and EDIT5, a T5-based approach tailored for editing Wikipedia articles with new evidence, highlighting the challenges and opportunities in neural generation models. Next, we will delve into PURR, the Petite Unsupervised Research and Revision model, designed to combat hallucinations in large language models. Through unsupervised training and efficient execution, PURR improves attribution and paves the way for robust editing methods. The talk will conclude with insights on the future of editing models and their transformative potential in managing and trusting information.
Dr. Forest Agostinelli
Artificial Intelligence Institute University of South Carolina
Forest Agostinelli is an assistant professor at the University of South Carolina. He received his B.S. from the Ohio State University, his M.S. from the University of Michigan, and his Ph.D. from the University of California, Irvine under Professor Pierre Baldi. His research aims to automate the discovery of new knowledge using deep learning, reinforcement learning, search, and formal logic. He looks to apply his research to fields such as puzzle solving, chemical synthesis, program synthesis, theorem proving, robotics, and education. His homepage is located at https://cse.sc.edu/~foresta/.
Title : Specifying Goals to Deep Neural Networks with Answer Set Programming
Abstract : Deep reinforcement learning has recently been used to train deep neural networks (DNNs) as heuristic functions to solve planning problems. While this approach has resulted in successfully solving planning problems with little to no domain-specific knowledge, specifying a goal to a DNN requires having prior knowledge about what states meet that goal specification and potentially requires re-training the DNN for that particular goal. To address this problem, we introduce a method of training a DNN to estimate the distance between a given state and a set of states, where a set of states is represented as a set of ground atoms in first-order logic. We then use first-order logic to specify goals and use answer set programming to obtain a set of ground atoms that represents that goal. As a result, one can invent new properties in first-order logic and need only specify what properties a goal state should or should not have without any need to know what states actually meet the specification. In our experiments with the Rubik's cube, we show that we can specify and reach goals without any need to re-train the DNN.
Dr. Pavan Kapanipathi
IBM RESEARCH, NY
Pavan Kapanipathi is a Principal Research Scientist at IBM Research. His primary research interests include Knowledge Graphs, Reasoning, Semantic Web, and Natural Language Processing. His works over the years have focused on leveraging knowledge graphs for reasoning tasks in various domains such as Knowledge Graphs, Neuro-Symbolic AI, Natural Language Processing, and Recommendation Systems. He is also the Reasoning PIC chair at IBM Research. Pavan Kapanipathi has over 50+ publications. His works have been published at Artificial Intelligence and Web (Semantic Web) conferences such as AAAI, WWW, ISWC, and ESWC. His recent work on complex question answering won the best paper award at MRQA workshop in ACL. He also has had a winning entry at the Triplification challenge (I-Semantics conference). He has co-organized Reasoning for Complex Question Answering Workshops at AAAI 19 and AAAI 20 and served as a PC member and reviewer for Artificial Intelligence (AAAI, IJCAI, NeurIPS), Computational Linguistics (EMNLP, ACL, NAACL), and Semantic Web conferences (ISWC, ESWC).
Title : A self-supervised rule learning framework for language understanding
Abstract : In the era of Large Language Models, there is still a need for interpretable and reliable language understanding. In pursuit of this, inspired by the masked language modeling setting of large language models, I will present our new neuro-symbolic framework for self-supervised rule learning from natural language. The framework is composed of three primary components (a) Semantic Parser to translate text to logic and logic to text; (b) A rule learner that takes in logic as input and output learnt rules based on self-supervision; (c) A reasoner that can reason over the rules and inform the semantic parser, the rule learner, and perform aspects of neuro-symbolic inference. In this talk, I will introduce the framework and go into the details of each of the above components. Our initial results have shown improved results in triple extraction tasks for semantic parsing and relation linking for the self-supervised rule learning work. The integration of all the three components of the framework is still in progress .
Dr. Mayur Naik
UNIVERSITY OF PENNSYLVANIA
Mayur Naik is a professor in the department of Computer and Information Science at the University of Pennsylvania. He is broadly interested in topics related to programming languages and artificial intelligence. His current research is motivated by the need to make AI applications safe, interpretable, data-efficient, and easier to develop. To this end, his research group is developing principled yet practical approaches to neurosymbolic programming, and applying them in high-stakes domains like healthcare and robotics. He obtained his Ph.D. in Computer Science from Stanford University in 2008, and was a research scientist at Intel Labs, Berkeley and a faculty at Georgia Tech.
Title : Scallop ( A Language for Neurosymbolic Programming )
Abstract : Neurosymbolic learning is an emerging paradigm which, at its core, combines the otherwise complementary worlds of classical algorithms and deep learning; in doing so, it ushers in more accurate, interpretable, and domain-aware solutions for today's most complex machine learning challenges. I will present Scallop, a general-purpose programming language that allows a wide range of modern AI applications to be written and trained in a data and compute efficient manner. I will illustrate how Scallop enables to specify algorithmic reasoning in diverse and challenging AI tasks, provides a succinct interface for machine learning programmers to integrate logical domain-specific knowledge, and outperforms state-of-the-art deep neural network models.
This is joint work with PhD students Ziyang Li and Jiani Huang.
