Seminar 3 : Marrying Reinforcement Learning and Deep Learning

The Computer Science Graduate Society is pleased to present the first talk of the winter semester by Emmanuel Bengio on 15th January. To help us know how much food to order, please fill this Google form if you plan to attend: http://goo.gl/forms/9lVPNT5tyV

—————–

Date: January 15
Time: 13:00-14:00
Location: McConnell Room 103

Speaker: Emmanuel Bengio
Title: Marrying Reinforcement Learning and Deep Learning

Abstract: In the last few years Machine Learning has boosted in popularity thanks to the outrageously successful applications of Deep Learning, in many areas including speech recognition, fraud detection, advertisement, recommendation systems, and, probably most famously, computer vision.
Taking advantage of these techniques, we have also very recently seen successful use of Deep Learning as a tool inside of Reinforcement Learning tasks, such as DeepMind’s popular Atari model. A new approach, which is the main concern of my research, consists in studying the opposite direction: using Reinforcement Learning to augment Deep Learning models. This takes the form of conditional computing, visual attention, memory mechanisms, and much more, which I will discuss during this talk.

Seminar 2 : Efficient Collaborations with Trust-Aware Robots

The Computer Science Graduate Society is pleased to present a talk by Anqi Xu on 2nd December (next Wednesday). To help us know how much food to order, please fill this Google form if you plan to attend: http://goo.gl/forms/hkof94CtJC.

Date: December 2 (next Wednesday)

Time: 12:00-13:00

Location: McConnell Room 103

Speaker: Anqi Xu

Title: Efficient Collaborations with Trust-Aware Robots

Abstract:

In this work, we give autonomous robot agents the ability to infer their human collaborator’s changing trust states, and consider how this signal can be capitalized to improve the efficiency of human-robot teams. This trust-aware robot framework incorporates advances in online human-robot trust modeling and interactive behavior adaptation for autonomous agents. We build upon these components by introducing the novel formulation of trust-induced conservative control. This enables the robot agent to momentarily alter its behaviors in response to the human’s trust losses, as an active means to mend damage to the team relationship. We present two end-to-end instantiations of trust-aware robots for distinct task domains of aerial terrain coverage and interactive autonomous driving. Our empirical assessments comprise of a large-scale controlled study, as well as field evaluations with a smart car platform. These assessments quantitatively demonstrate the diverse efficiency gains of trust-aware robots.

Seminar 1 : RNA Sequence Design

newanigif

Speaker: Vladimir Reinharz
Title: RNA Sequence Design

Abstract:
RNAs are biomolecules that can be seen as words in {ACGU}+. An interesting mathematical representation of the structure as a dot bracket sequence [e.g. (((…))) ] has been shown to have some ressemblance with reality. Having an energy model derived from experiments, the ensemble of structures given a sequence has been completely characterized in the 70s (in this model). At the opposite, given a structure, knowing if it exists a sequence with an affinity higher than some threshold  is believed to be NP-hard. I developed an algorithm based on the inside-outside to, given a structure, do a global stochastic Boltzmann sampling of sequences in linear time. I will also talk of a few potential applications in drug design and bioengineering. It has been implemented in a tool called incaRNAtion

http://dx.doi.org/10.1093/bioinformatics/btt217