Schedule: November 23 (Thu), 2017
Venue: Room Z201-202, Graduate School of Science, Kobe University, Japan
Speakers
Program and Abstracts (PDF file)
Supported by
Contact
Cai Jianfei (Nanyang Technological University, Singapore)
Title: Learning-based 3D Visual Recognition and Reconstruction
Abstract: With the great success of applying deep learning in image classification,
object detection and semantic segmentation, deep learning has also been
applied for 3D visual recognition and computing. However, unlike 2D images,
learning-based 3D visual computing has its unique challenges such as diverse
data formats, distinct data modality, large data volume and lack of large-scale
well-labelled data. In this talk, I will introduce some of our recent works
on learning-based 3D visual computing including multi-model feature fusion
frameworks for RGB-D object recognition, RGB-D scene classification and
our very recent work on learning based 3D face reconstruction.
Title: Recent Challenges to Cybersecurity and Privacy-Preserving Data Mining
Using Machine Learning
Abstract: This presentation gives a brief introduction to our machine learning approaches
to cybersecurity and privacy-preserving data mining (PPDM). A mission of
our current cybersecurity projects is to construct a large-scale monitoring
system to detect cyberattacks on the Internet through Darknet Traffic Analysis,
and to construct an AI crawling system to monitor cybersecurity products
dark marketplaces on the Tor hidden service. I briefly show an interesting
observation of the darknet traffic before the source code of a notorious
IoT malware called ‘Mirai’ was first opened in September 2016. In our darknet
analysis, the frequent pattern mining was performed to a large set of TCP
SYN packets collected from July 1st 2016 to September 30th 2016 with the
NICT /16 darknet sensor. On the other hand, the purpose of our PPDDM project
is to provide a secure cloud outsourcing platform for client users who
requests an analysist to sensitive data including privacy information.
Currently, we are developing a Privacy Preserving Extreme Learning Machine
(PP-ELM) using additively homomorphic encryption, which supports summation
on encrypted data. In order to handle the learning over encrypted data
efficiently, we propose a three participants model as an implementation
of PP-ELM; data contributors, an outsourced server, and a data analyst.
We show a simple experimental result using a relatively small benchmark
dataset for the performance evaluation.
Title: Learning Deep Convolutional Networks for Visual Scene Understanding
Abstract: Automated understanding and analyzing the content of scene images and
videos is a fundamental and challenging problem for computer vision research, which is
the essential component in many emerging applications including intelligent video
surveillance, autonomous robots or vehicles, content based image retrieval, aerial image
analysis in remote sensing, just to name a few. Recently deep Convolutional Neural Nets
(CNNs) have demonstrated outstanding performance in semantic segmentation which is a
core task towards visual scene understanding. In this talk, I will present our recent CNN
based methods for accurate semantic segmentation. We exploit representations at
multiple levels of abstraction for high-resolution output and employ residual connections
with identity mappings for all network components, such that gradients can be directly
propagated through short-range and long-range residual connections allowing for both
effective and efficient end-to-end training.
Title: Extracting Breeding Cows from Monitored Camera Image Data
Abstract: The extraction of knowledge from large amount of data plays an important
role in improving productivity in recent agriculture. In the livestock field, methods for analyzing various data
(e.g, growth and body condition data), which are collected from an individual livestock,
have been developed to improve the management efficiency.
Our project team has been working on the exploration of measurement and analysis method to extract
information which contributes to raise productivity of breeding cows, and the development of innovative growth management
technology by examining the interaction of breeding cows. On this occasion, I would like to introduce
two recent topics of our research: (i) estimation of calf weight from fixed-point stereo camera images using
three-dimensional successive cylindrical model, and (ii) detecting and tracking breeding cows from bird's eye video of pasture.
Our goal is to grasp the growth status, health and stress conditions of cows by integrating multiple
technologies such as image processing, GPS, and various sensory data.
Title: Learning Market Parameters using Aggregate Demand Queries
Abstract: We study efficient algorithms for a natural learning problem in markets.
There is one seller with m divisible goods and n buyers with unknown individual utility functions and budgets of money. The seller can repeatedly announce prices and observe aggregate demand bundles requested by the buyers. The goal of the seller is to learn the utility functions and budgets of the buyers. Our scenario falls into the classic domain of “revealed preference” analysis. Problems with revealed preference have recently started to attract increased interest in computer science due to their fundamental nature in understanding customer behavior in electronic markets. The goal of revealed preference analysis is to observe rational agent behavior, to explain it using a suitable model for the utility functions, and to predict future agent behavior. Our results are the first polynomialtime algorithms to learn utility and budget parameters via revealed preference queries in classic Fisher markets with multiple buyers. Our analysis concentrates on linear, CES, and Leontief markets, which are the most prominent classes studied in the literature. Some of our results extend to general Arrow-Debreu exchange markets.
Title: Voice Conversion Based on Non-negative Matrix Factorization
Abstract: Voice conversion (VC) is a technique for converting specific information in speech,
while preserving the other information in the utterance. The most popular VC application is speaker conversion,
which converts a source speaker's voice individuality to that of a specified target speaker,
while preserving the linguistic information. VC has also been applied to emotion conversion.
In recent years, exemplar-based VC, which is a non-statistical approach, has been attracting interest.
In this talk, I will introduce exemplar-based VC using non-negative matrix factorization that has some advantages
to noise-robustness, small-parallel corpus (training data), many-to-many (arbitrary speakers) VC, and multi-modal VC.
Title: High-Dimensional Portfolio Selection with DECODE
Abstract: This paper investigates the high-dimensional portfolio selection problem in
which the number of risky assets is greater than the number of observation times. It is
well known that the theoretically optimal portfolios, subject to estimation errors, perform
poorly in many empirical studies, especially in the presence of lots of assets. This paper
proposes a novel statistical learning framework, called Descent-based Calibrated Optimal
Direct Estimation (DECODE), which is free of tuning parameters and directly estimates
effective parameters appearing in the optimal portfolios. The resulting DECODE portfolios
are sparse, which realize data-driven selection of favourable assets. The advantages of
the DECODE approach also include its computational superiority (much faster than those
with cross-validation) and its applicability for high-dimensional cases and non-Gaussian
distributions of asset returns. This paper proves the consistency results for the DECODE
approach. Numerical and empirical studies are conducted to compare the performances
of DECODE and other existing competitor schemes. If time permits, I will briefly discuss
about the further applications of DECODE in Finance.
Title:Wearable Sensing and Information Presentation Considering Psychological Effects
Abstract: We have developed many techniques of sensing and information presentation in
wearable computing environments, to improve the quality of life and the quality of professional works.