[Adapt] 7月10日(今天)9:00-12:00--陈露、李松泽、洪义博士

Kenny Zhu kzhu at cs.sjtu.edu.cn
Fri Jul 10 08:55:51 CST 2020




会议号:686 465 68231


参会链接: <https://zoom.com.cn/j/68646568231>




Title: Universal Conversational Agents with Structured Deep Learning


Abstract:A task-oriented spoken dialogue system is a system that can
continuously interact with human to accomplish a predefined task through
speech. Dialogue policy plays an important role in task-oriented spoken
dialogue systems. It determines how to respond to users. Deep reinforcement
learning (DRL) approaches have been widely used for policy optimization.
However, these deep models are still challenging for two reasons: first,
many DRL-based policies are not sample-efficient; and second, most models do
not have the capability of policy transfer between different domains. In
this talk, I will introduce our proposed universal framework, AgentGraph, to
tackle these two problems. The model AgentGraph is the combination of graph
neural network (GNN) based architecture and DRL-based algorithm. It can be
regarded as one of the multi-agent reinforcement learning approaches. In
addition to the universal dialogue policy AgentGraph, I will also briefly
introduce our proposed universal dialogue state tracking models.


Bio: Lu Chen obtained his Ph.D. degree in computer science and engineering
from Shanghai Jiao Tong University (SJTU) in 2020. Before coming to SJTU, he
received a Bachelor's degree in computer science and engineering from
Huazhong University of Science & Technology (HUST) in 2013. His research
interests include dialogue systems, reinforcement learning, structured deep
learning. The goal of his research is to build evolvable and universal
conversational agents, which can converse with humans among many domains and
improve their performance with various signals. He has authored/co-authored
more than 20 journal articles (e.g. IEEE/ACM transactions) and peer-reviewed
conference papers (e.g. ACL, EMNLP, AAAI, COLING, ICASSP), one of them was
selected as COLING2018 Area Chair Favorites.





Title: Nakamoto meets Shannon: Scaling blockchains using codes


Abstract:Current blockchain systems do not scale with the network resources.
Sharding, as a promising proposal to achieve horizontal scalability, fail to
maintain security of the system in the presence of adaptive adversaries.
This talk will cover an emerging paradigm of designing scalable and secure
blockchain protocols using error correcting codes. Specifically, we consider
the following two problems 1) transaction verification, and 2) verifying
data availability for light clients. For transaction verification, we
propose PolyShard that creates coded shard ledgers and input transactions
using Lagrange polynomial interpolation. PolyShard scales system throughput
with the network size, while maintaining the security of verification
results via Reed-Solomon decoding. For the problem of verifying data
availability, we propose a novel cryptographic accumulator named Coded
Merkle Tree (CMT) to commit a block. CMT iteratively encodes a block and the
hashes of its chunks using a family of regular LDPC codes, into a constant
number of hash values that are stored in the block header. CMT enables
verifying the availability of a block with constant number of samples, a
liner decoding complexity, and a constant size of fraud proof for incorrect


Bio: Dr. Songze Li is a research scientist at Stanford University working on
designing and developing next-generation scalable blockchain systems. Prior
to that, Dr. Li spent a year as a research scientist with Applied Protocol
Research, where he invented a novel cryptographic accumulator Coded Merkle
Tree, which enables efficient verification of data availability for light
clients in blockchains. Songze received his Ph.D. degree from University of
Southern Californi <http://www.usc.edu/> a in 2018, and his B.Sc. degree
from New York <http://engineering.nyu.edu/>  University in 2011, both in
electrical engineering. Songze’s research interests lie on the intersection
of theory and system of designing efficient, scalable, and secure
distributed computing solutions for machine learning and blockchains.
Specifically, he first introduced leveraging techniques from
information/coding theory to design distributed computing algorithms, which
opened up a new research direction of code design for speeding up
computations. Songze received USC Viterbi School of Engineering Doctoral
Fellowship in 2011. He is among Qualcomm Innovation Fellowship Finalists in




Title: Parametric Regression for Medical Analysis: Beyond Euclidean Space



Uncovering time-varying trends from spatiotemporal data is important in
medical analysis, for instance, in understanding brain development, aging,
and disease progression. Data for these studies have different types, e.g.,
shapes, image scans, videos, which have complex structures and are best
treated as elements of non-Euclidean spaces. In this talk, I present
parametric regression models on two non-Euclidean spaces, i.e., the
Grassmann manifold and the manifold of diffeomorphisms, to handle multiple
types of spatiotemporal data. These models are generalized from Euclidean
regression and provide efficient and straightforward solutions for analyzing
non-Euclidean data. The experimental results demonstrate the resulting
models’ effectiveness in studying Alzheimer’s Disease. 



Dr. Yi Hong is a tenure-track assistant professor of CS at the University of
Georgia (UGA). Before joining UGA, she received her Ph.D. from the
University of North Carolina (UNC) at Chapel Hill in 2016. Dr. Hong's
primary research interest lies in developing novel techniques for medical
multimedia data, mainly in medical image understanding, statistical shape
analysis, and population studies. Dr. Hong is a recipient of the 2014 MICCAI
young scientist award, a 2015-2016 UNC dissertation completion fellowship,
and two NSF grant awards.








chen-zx at cs.sjtu.edu.cn


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs.sjtu.edu.cn/pipermail/adapt/attachments/20200710/42dee6f8/attachment-0001.html>

More information about the Adapt mailing list