[cs-talks] Upcoming CS Seminars: IVC Seminar (Tues) + Theory Sem (Fri)
fgreen1 at bu.edu
Mon Nov 9 15:51:35 EST 2015
Object Detectors Emerge in the Convolutional Neural Networks Trained for Scene Recognition
Bolei Zhou, MIT CSAIL
Tuesday, November 10, 2015 at 2pm in MCS 148
Abstract: With the success of new computational architectures for visual processing, such as convolutional neural networks (CNN) and access to image databases with millions of labeled example (e.g., ImageNet, Places), the state of the art in computer vision is advancing
rapidly. One important factor for continued progress is to understand the representations that are learned by the inner layers of these deep learning architectures. Here we show that object detectors emerge in the CNNs trained to perform scene recognition on Places Database. As scenes are composed of objects, the CNN for scene classification automatically discovers meaningful objects detectors, representative of the learned scene categories. With object detectors emerging as a result of learning to recognize scenes, our work demonstrates that the same network can perform both scene recognition and object localization in a single forward-pass, without ever having been explicitly taught the notion of objects. The project page of Places Database and Places CNNs is at thttp://places.csail.mit.edu<http://places.csail.mit.edu/>
Recent Work on the Global Network Alignment Problem
Ben Hescott, Tufts
Friday, November 13, 2015 at 12pm in MCS 135
Abstract: In this work we shift focus in the global network alignment problem, moving away from identifying local structural similarities, and focusing instead on finding coherent, functionally related groups of genes across species. We introduce a new solution, CANDL a Coarsely Aligning Networks with Diffusion and Landmarks. Key to this technique is a new method for embedding the graph into a continuous metric space. Unlike previous methods that seek to conserve local motifs, this technique identifies neighborhoods that are functionally similar. In the second part of the talk, we identify and quantify previously overlooked limitations of topological validation techniques for network alignment problem. We show that all current techniques for comparing the quality of aligners fall short and that functional techniques are required.
Bio: Benjamin Hescott is a Assistant Professor in the Department of Computer Science at Tufts University's School of Engineering. His research interests include computational complexity, approximation algorithms, and computational biology. He graduated from Boston University with a Ph.D. in computer science in 2008. He is the faculty supervisor for the student ACM chapter and serves as liaison to the New England Undergraduate Computer Science Symposium. He is member of the leadership team for ELA (Empowering Leadership Alliance) whose main purpose is encouraging, preparing, and retaining underrepresented minorities in computer science.
Ben is the recipient of the 2011 IEEE Computer Society Computer Science and Engineering Undergraduate Teaching Award, the 2011 Lerman-Neubauer Prize for Outstanding Teaching and Advising, the 2012 Henry and Madeline Fischer Award (Engineering Teacher of the year award),
the 2012 Lillian and Joseph Leibner Award for Excellence in Teaching and Advising of Students and the 2013 Tufts Graduate Student Council Award for Outstanding Faculty Contribution to Graduate Studies.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cs-talks