Скачать с ютуб Yilun Du - Implicit Learning with Energy-Based Models | Nuro Technical Talks в хорошем качестве

Yilun Du - Implicit Learning with Energy-Based Models | Nuro Technical Talks

About the Talk: Deep learning has performed well on internet datasets, but still faces challenges when applied to complex embodied domains. In this talk, I’ll discuss how energy based models are a useful tool for constructing deep learning systems in embodied domains. First, I will illustrate how energy based optimization enables us to compose multiple different models together, enabling us to compositionally understand different visual scenes. Next, I will illustrate how an energy based optimization procedure enables us to generate diverse behaviors with respect to test time constraints, enabling us to adapt, at test time, to different goals and rewards. Finally, I’ll talk about how energy optimization enables us to iteratively reason and adapt to novel problems, leveraging more complex optimization procedures to solve more complex algorithmic problems at test time. About the Speaker: Yilun Du is a fourth year graduate student at MIT EECS. Previously, he has done a research fellowship at OpenAI, and research internships at FAIR, DeepMind and Google Brain. He is supported by an NSF Graduate Fellowship and was a finalist for Qualcomm and Open Philanthropy fellowships.