34
@ U S T . H K
REINFORCEMENT
LEARNING
DEEP LEARNING
TRANSFER LEARNING
FromNumbers to Knowledge
For a person, the ability to recognize
and apply knowledge and skills learned
in previous tasks to new endeavors is a
natural occurrence. After understanding
how one card game works, it is easier
to pick up another. For a computer,
such learning is incredibly hard.
This is the specialist domain of Prof
Qiang Yang, an expert in data
mining, artificial intelligence, machine
learning, transfer learning and deep
reinforcement learning.
Prof Yang has spent 20 years
fathoming algorithms that seek to endow
computers with similar capabilities
to humans in retaining and reusing
previously learned knowledge in order
to “think” and “decide” how to extract
information and patterns from the rivers
of data flooding our digital age. Prof Yang
and his team have improved the accuracy
of computers’ performance through
devising versatile frameworks for such
“transfer learning”. He has developed
Instance-based Transfer Learning, which
uses individual instances
BUILDING
INTELLIGENCE
to bridge different domains, and
Heterogeneous Transfer Learning, where
the computer learns in one knowledge
domain (for example, text) then transfers
what it has learned to a separate or more
difficult domain (for example, images).
Prof Yang has made these frameworks
open source, enabling other researchers
and the field overall to develop at a faster
pace. He was also the first to propose the
use of transfer learning in collaborative
filtering and recommender systems.
Applications have ranged from early
online advertising directed at users
to improvements in recommendation
systems, including a state-of-the-art
recommendation system for ICT global
giant Huawei’s App Store.
Recent research at the WeChat-
HKUST Joint Lab on Artificial Intelligence
Technology (WHAT LAB), set up with
Mainland China internet giant Tencent
in 2015, has inspired a novel application
to improve machine reading capabilities.
Books, news articles, and blogs are used
as input to train a machine learning
model that can produce an
abstract of such readability that it
doesn’t appear to have been written by a
machine. The objective is to assist people
with information overload on social
networks or boost company productivity
by enabling a computer to develop an
abstract of a long report or integrate data
and highlight the main points the reader
needs to know. By reading books by the
same author, the demonstration model
designed by Prof Yang and his students
has even written a high-quality novel of
its own in the writer’s style, taking just a
few seconds to do so.
In improving such machine reading
abilities, Prof Yang’s team has
become the first to integrate a
reinforcement learning algorithm
that leverages users’ feedback
related to positive comments
on prior abstracts with transfer
learning and deep learning
(recurrent neural networks) to
help the computer make a more
intelligent decision on what abstract
to generate. The innovation has
improved the quality substantially.
With information to hand quicker, it
can also speed up report-writing as
well as learning.
“The work at HKUST seeks to
increase the knowledge we can get
from data by making the process
of moving from data source to
understanding faster and more
efficient, accurate and useful to
people,” Prof Yang said.
Cycling
Motorcycling
Humans have the ability
to apply knowledge and
skills learned in previous
tasks (e.g. cycling) to new
tasks (e.g. motorcycling).
Prof Yang’s transfer
learning algorithms help
computers to acquire
the ability to retain and
transfer knowledge from
a source domain to
a target domain.
We are
inventors,
always
thinking of
how to use data
in a new way
PROF QIANG YANG
New Bright Professor of Engineering,
Head, Department of Computer Science
and Engineering,
Director, HKUST Big Data Institute,
Inaugural Editor-in-Chief of
IEEE Transactions on Big Data
Prof Yang combines three levels of machine learning techniques
to achieve highly innovative machine reading capabilities.
