Ian Goodfellow

Title: Defense Against the Dark Arts: Making Machine Learning Robust to Adversarial Examples

Abstract: Machine learning models are easily deceived by adversarial examples: inputs that are optimized to produce incorrect responses. Adversarial examples for a target model may be constructed without access to the model or even without knowing which algorithm was used to train the model. Adversarial examples for computer vision models fool the model even when displayed in the physical world and observed from multiple viewpoints. In this talk, I'll describe the best known defenses against adversarial examples.

 Ruslan Salakhutdinov

Title: Neural Map: Structured Memory for Deep Reinforcement Learning

Abstract: A critical component to enabling intelligent reasoning in partially observable environments is memory. However, Deep Reinforcement Learning (DRL) agents have so far used relatively simple memory architectures, with the main methods to overcome partial observability being either a temporal convolution over the past k frames or an LSTM layer. In this talk, we will introduce a memory system with an adaptable write operator that is customized to the 3D environments that DRL agents typically interact with. The architecture, called the Neural Map, uses a spatially structured 2D memory image to learn to store arbitrary information about the environment over long time lags. We demonstrate empirically that the Neural Map surpasses previous DRL memories on a set of challenging 2D and 3D maze environments and show that it is capable of generalizing to environments that were not seen during training. Joint work with Emilio Parisotto
 Joaquin Quiñonero Candela

Title: Lessons from powering Facebook experiences at scale with AI

Abstract: At Facebook, we are building AI that helps bring the world closer together. Our work in areas like computer vision, natural language understanding, speech and core machine learning will help people better understand each other and the world. However, how does one enable a very large number of engineers with varying levels of AI expertise to explore as fast as possible the entire funnel, from devising new modeling paradigms all the way to rapid live experimentation, and ultimately to product deployment at Facebook scale?  In this talk I will give an overview of the different applications of AI across the entire family of Facebook experiences, and share some of the lessons we have learnt, both from the perspective of tools and infrastructure, as well as from the lens of culture and organizational principles.
 Adam Gazzaley
UCSF, Neuroscape

Title: Technology meets Neuroscience - A Vision of the Future of Brain Optimization

Abstract: A fundamental challenge of modern society is the development of effective approaches to enhance brain function and cognition in both the healthy and impaired. For the healthy, this should be a core mission of our educational system and for the cognitively impaired this is the primary goal of our medical system. Unfortunately, neither of these systems have effectively met this challenge. I will describe a novel approach out of UCSF Neuroscape that uses custom-designed video games to achieve meaningful and sustainable cognitive enhancement via personalized closed-loop systems (Nature 2013; Neuron 4014). I will also share with you the next stage of our research program, which integrates our video games with the latest technological innovations in software (e.g., brain computer interface algorithms, GPU computing, cloud-based analytics, machine learning) and hardware (e.g., virtual reality, mobile EEG, motion capture, physiological recording devices, transcranial brain stimulation) to further enhance our brain’s information processing systems with the ultimate aim of improving quality of life. I will highlight opportunities to use machine learning approaches to create an integrated multimodal closed-loop system.