[cs-talks] CS Upcoming Seminars: Hariri-Guest Speaker (Wed) + IVC (Thurs)
fgreen1 at bu.edu
Wed Mar 16 11:09:11 EDT 2016
An Optical Turing Machine for Network Processing
(with background on a digital optical Internet router)
Joe Touch, USC/ISI
Wednesday, March 16, 2016 at 1:30pm in MCS 180- Hariri Seminar Room
Abstract: The Optical Turing Machine (OTM) is an approach to digital optical processing that supports computation in the same format used for high-speed transmission. This talk identifies the key capabilities required to support native digital optical processing for typical in-network functions including forwarding, security, and filtering. Current analog and binary digital approaches – including optical transistors – are considered and shown insufficient for optical networks. The requirements for developing optical communication and computation using a single encoding are presented, as are the capabilities required for network computation. Recent results in regenerating N-PSK signals using non-degenerate PSA and multilevel amplitude squeezing are presented, with an analysis of several alternate approaches and their compatibility with optical computation. OTM was motivated by the need to support optical Internet routing, and this talk also presents the design for such a router based on decomposing the steps required for IP packet forwarding. Implementations of hop-count decrement and header matching are coupled with a recent simulation-based approach to variable-length packet merging that avoids recirculation, resulting in an all-optical data plane. A method for IPv4 checksum computation is presented and the implications of this design are
considered, including the potential for chip and system integration.
Bilinear Models for Fine-grained Visual Recognition
Subhransu Maji, UMass Amherst
Thursday, March 17, 2016 at 2pm in MCS 148
Abstract: Fine-grained recognition tasks such as distinguishing “California Gulls” from “Ringed-beak Gulls” are challenging because the subtle differences between categories are often confounded by factors such as pose, viewpoint, and background variation. Currently there are two classes of models that work well for such tasks. First are part-based which proceed by localizing parts followed by a more detailed analysis. The second are variants of bag-of-visual-words models that were originally developed for texture recognition. The part-based methods are more accurate but need extra annotations at training time limiting their scalability.
I’ll introduce bilinear models which are particularly suitable for modeling localized appearance variations common in fine-grained recognition tasks. It consists of two feature extractors whose outputs are multiplied using outer product at each location of the image and pooled to obtain a descriptor. We show that various texture descriptors such as Fisher-vector and VLAD can be written as bilinear models. The architecture also resembles the computations of part-based models used for fine-grained tasks. Moreover, when these feature extractors are based on CNNs, the bilinear models can be trained end-to-end using image labels only, and remarkably outperforms several fully supervised part-based CNN models on a number of fine-grained datasets. For instance we obtain 84.1% accuracy on the CUB-200-2011 dataset requiring no part or bounding-box annotations during training or testing. I’ll also present the nature of invariances captured by these models though visualizations.
Bio: Subhransu Maji is an Assistant Professor in the College of Information and Computer Sciences at the University of Massachusetts, Amherst. Previously, he was a Research Assistant Professor at TTI Chicago, a philanthropically endowed academic computer science institute in the University of Chicago campus. He obtained his Ph.D. from the University of California, Berkeley in 2011, and B.Tech. in Computer Science and Engineering from IIT Kanpur in 2006. His research focusses on developing visual recognition architectures with an eye for efficiency and accuracy.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cs-talks