Salesforce Tech Talk: Towards Versatile AI: Multi-task Learning and Generalization to New Tasks
Speaker
Stephan Zheng
Salesforce Research
Host
Callie Mathews
CSAIL Alliances
Please register or email Callie Mathews cmathews@csail.mit.edu to confirm attendance
https://www.eventbrite.com/e/salesforce-tech-talk-towards-versatile-ai-multi-task-learning-and-generalization-to-new-tasks-tickets-59182314897
April 23rd, 12PM-1PM StarD463
Lunch will be provided
Salesforce Tech Talk: Towards Versatile AI: Multi-task Learning and Generalization to New Tasks
Deep neural networks have been top performers for machine learning problems on a single task, such as machine translation or playing Atari video-games. However, it is hard to achieve strong performance on multiple tasks simultaneously or generalization to unseen tasks, as data statistics and performance metrics can vary dramatically across tasks. In this talk, I will present two recent works that address these challenges.
First, I will show decaNLP, a framework and competition for multi-task learning that unifies ten natural language processing (NLP) tasks as question-answering problems. This approach enables training unified models that can achieve competitive performance simultaneously on the ten NLP tasks, which include translation, summarization and text classification.
Second, I will present how to derive theoretical generalization guarantees in reparametrizable reinforcement learning, in which trajectory distributions can be decomposed using the reparametrization trick. We theoretically derive and empirically verify Rademacher/PAC-Bayes generalization bounds for both intrinsic (due to overfitting within a single task) and external errors (due to shifts in world dynamics between tasks).
Finally, I will give a high-level overview of machine learning research at Salesforce Research, including research on AI for Social Good and explainable AI.
Presenter: Stephan Zheng, Research Scientist Salesforce Research
Bio:
Stephan Zheng is a Research Scientist at Salesforce Research. He obtained his PhD in 2018 in the Machine Learning group at Caltech, advised by Professor Yisong Yue. His current research focuses on deep reinforcement learning in multi-agent environments. He has also worked on improving the robustness of deep learning and multi-resolution learning for spatiotemporal data.
Previously, he received an MSc (Theoretical Physics) and BSc (Physics, Mathematics) from Utrecht University, read Part III Mathematics at the University of Cambridge and was a visiting student at Harvard University. He received the 2011 Lorenz Prize in Theoretical Physics from the Dutch Academy of Arts and Sciences, and was twice a research intern with Google Research and Google Brain.
https://www.eventbrite.com/e/salesforce-tech-talk-towards-versatile-ai-multi-task-learning-and-generalization-to-new-tasks-tickets-59182314897
April 23rd, 12PM-1PM StarD463
Lunch will be provided
Salesforce Tech Talk: Towards Versatile AI: Multi-task Learning and Generalization to New Tasks
Deep neural networks have been top performers for machine learning problems on a single task, such as machine translation or playing Atari video-games. However, it is hard to achieve strong performance on multiple tasks simultaneously or generalization to unseen tasks, as data statistics and performance metrics can vary dramatically across tasks. In this talk, I will present two recent works that address these challenges.
First, I will show decaNLP, a framework and competition for multi-task learning that unifies ten natural language processing (NLP) tasks as question-answering problems. This approach enables training unified models that can achieve competitive performance simultaneously on the ten NLP tasks, which include translation, summarization and text classification.
Second, I will present how to derive theoretical generalization guarantees in reparametrizable reinforcement learning, in which trajectory distributions can be decomposed using the reparametrization trick. We theoretically derive and empirically verify Rademacher/PAC-Bayes generalization bounds for both intrinsic (due to overfitting within a single task) and external errors (due to shifts in world dynamics between tasks).
Finally, I will give a high-level overview of machine learning research at Salesforce Research, including research on AI for Social Good and explainable AI.
Presenter: Stephan Zheng, Research Scientist Salesforce Research
Bio:
Stephan Zheng is a Research Scientist at Salesforce Research. He obtained his PhD in 2018 in the Machine Learning group at Caltech, advised by Professor Yisong Yue. His current research focuses on deep reinforcement learning in multi-agent environments. He has also worked on improving the robustness of deep learning and multi-resolution learning for spatiotemporal data.
Previously, he received an MSc (Theoretical Physics) and BSc (Physics, Mathematics) from Utrecht University, read Part III Mathematics at the University of Cambridge and was a visiting student at Harvard University. He received the 2011 Lorenz Prize in Theoretical Physics from the Dutch Academy of Arts and Sciences, and was twice a research intern with Google Research and Google Brain.