CS Colloquium

Have you ever wondered how Siri, Alexa, or any chat-bot perceives you? We are interested in answering this question by making it possible to receive feedback from interactions with AI. Understanding the human message and emotion when approaching a machine is a valuable psychological experience and can tell us a lot about ourselves and our society. In collaboration with the Art Department, we are using AI methods to create a reaction to people’s interaction with machines.  The interaction received through sensory inputs is processed by the machine and the content of the interaction is decoded. Using deep learning models for natural language processing, the emotional state of the interaction is detected from the content. The detected state is used to visualize, in 2D or 3D, what the machine has perceived from the interaction. The final product will give the user a personalized interaction with the machine through a piece of art that reflects the user's emotion. This product can be combined with VR equipment and used for providing more realistic experience for game players, online training participants, or interaction with chat-bots.

HPC underwent several dramatic changes in system architecture over the last 50 years. From the pre-70’s mainframe, to vector processors and multiprocessors of the 80s, followed by the emergence of the “killer micros” and shift to commodity processors and clusters. From shared memory to distributed memory systems, and the addition of accelerators (GPUs). This journey is accompanied by tracking the transition from a single thread program execution to ever increasing levels of parallelism, and the implications to the software tools and the application end user. HPC today is much more than the domain of numerical simulations. It includes data analytics and AI.

Dr. Cooper will describe some of the challenges involved in building a large-scale distributed cache invalidation system and will present the design for one such system, called Thialfi, which was built and operated at Google for most of the past decade. He will also discuss the limitations of that design and touch on ways in which modern infrastructure allows improvements to it. The work is joint with Atul Adya, Phil Bogle, Dennis Geels, Brice Hulse, Larry Kai, Vishesh Khemani, Nick Kline, Colin Meek, Amanda Moreton, Daniel Myers, and Michael Piatek.

The blind or visually impaired often miss out on the visual information conveyed through videos. The vast majority of online video material is currently not accessible to millions of visually-impaired people who would significantly benefit from improved access to videos for education, employment, and entertainment purposes.

This work addresses two major issues:

1. Enhancing video accessibility for blind or visually-impaired individuals.
2. Generating well-structured training data to advance the state of the art in video understanding.

Zoox is transforming mobility-as-a-service by developing a fully autonomous, purpose-built fleet that is designed for AI to drive and humans to enjoy. In the first part of this presentation I will introduce the main components of the Zoox's software stack and its hardware. I will discuss the challenges of localization, perception, prediction, and planning in autonomous vehicles. In the second part of this presentation, I will share some resources that students can use to learn more about different aspects of autonomous vehicle technology and prepare to become a researcher or practitioner in this field.