Explainable Deep Learning for Natural Language Processing - Mihai Surdeanu
Abstract:
While deep learning approaches to natural language processing (NLP) have had many successes, they can be difficult to understand, augment, or maintain as needs shift. In this talk I will discuss two recent efforts that aim to bring explainability back into deep learning methods for NLP.
In the first part of the talk, I will introduce an explainable approach for information extraction (IE), an important NLP task, which mitigates the tension between generalization and explainability by jointly training for the two goals. Our approach uses a multi-task learning architecture, which jointly trains a classifier for information extraction, and a sequence model that labels words in the context of the relation that explain the decisions of the relation classifier. This sequence model is trained using a hybrid strategy: supervised, when supervision from pre-existing patterns is available, and semi-supervised otherwise. In the latter situation, we treat the sequence model's labels as latent variables, and learn the best assignment that maximizes the performance of the extractor. We show that, even with minimal guidance for what makes a good explanation, i.e., 5 rules per relation type to be extracted, the sequence model provides labels that serve as accurate explanations. Further, we show that the joint training generally improves the performance of the IE classifier.
In the second part of the talk, we adapt recent advances from the adjacent field of program synthesis to information extraction, synthesizing extraction rules directly from a few provided examples. We use a transformer-based architecture to guide an enumerative search, and show that this reduces the number of steps that need to be explored before a rule is found. Further, we show that without training the synthesis algorithm on the specific domain, our synthesized rules achieve state-of-the-art performance in a 1-shot IE task, i.e., when only 1 example is provided for each class to be learned.
Bio:
Dr. Surdeanu earned a Ph.D. in Computer Science from Southern Methodist University, Dallas, Texas, in 2001. He has more than 15 years of experience in building systems driven by natural language processing (NLP) and machine learning. His experience spans both academia (Stanford University, University of Arizona) and industry (Yahoo! Research and two NLP-centric startups). During his career he published more than 80 peer-reviewed articles, including two articles that were among the top three most cited articles at two different NLP conferences. He was a leader or member of teams that ranked in the top three at seven highly competitive international evaluations of end-user NLP systems such as question answering and information extraction. His work was funded by several government organizations (DARPA, NIH), as well as private foundations (the Allen Institute for Artificial Intelligence, the Bill & Melinda Gates Foundation). Dr. Surdeanu's current work focuses on using machine reading to extract structure from free text, and using this structure to construct causal models that can be used to understand, explain, and predict hypotheses for precision medicine.
A Safety View on Generalization for Machine Learning - Alexandru Paul Condurache
Abstract:
As the practical footprint of machine learning (ML) constantly enlarges to include even more new application areas, the topic of safety becomes of major concern. Traditional approaches to safety leverage causality. However, due to the correlation-based nature of the currently dominating ML methods, a new take on safety is needed. In this context, we need to answer in a convincing manner the same key question of finding out the root causes of a failure. Generalization is the ability to correctly decide on previously unseen data. Optimizing the generalization ability, which lies at the heart of ML, clearly implies dealing with generalization failures and is therefore inherently related to safety. In this talk, I will discuss the interplay between ML safety and generalization with a focus on leveraging prior knowledge as one of the key considerations of a safety argumentation for ML.
Bio:
Alexandru Paul Condurache received the Dipl.-Ing. degree in electrical engineering from the 'Politehnica' University of Bucharest, Romania in 2000, being awarded the Werner von Siemens Excellence Award for his diploma thesis. He obtained the Diploma of Advanced Studies in biomedical engineering also from the 'Politehnica' University of Bucharest, Romania, in 2001. In 2007 he received the Dr.-Ing. degree in computer science from the University of Lübeck, Germany under the supervision of Prof. Til Aach. In 2014 he habiltated in computer science at the University of Lübeck. Between 2002 and 2013, he was with the Institute for Signal Processing, University of Lübeck. Until 2007 he was a doctoral student, specializing in medical image analysis. Afterwards he conducted his postdoctoral research in information forensics (biometric authentication, surveillance and event detection) and discriminant feature analysis with a focus on leveraging a-priori information to improve the performance of machine-learning solutions. He is now with Robert Bosch GmbH where he contributes towards an Artificial Intelligence based approach to Autonomous Driving.