40
@ U S T . H K
Robots now coming into being in HKUST labs are
like babies. A child initially knows nothing of the
world. But using its five senses – hearing, sight,
touch, taste, and smell – an infant has the capacity
to rapidly learn from what it perceives in the
immediate environment and how that environment
changes in response to its actions.
This is the perception-action cycle, understanding
that is also vital for the development of the next
generation of robots. The new robot may start its
existence as a blank slate. But if, like a child, it has the
capacity to teach and correct itself, it can do much
more than a pre-programmedmachine.
Prof Bertram Shi is leading research where
electronics and computing meet biology and
neuroscience. Prof Shi is renowned in the robotics
community for his work in developing the Active
Efficient Coding (AEC) framework for machine
learning in collaboration with Prof Jochen Triesch’s
team at the Frankfurt Institute for Advanced Studies
in Germany. This framework extends Horace
Barlow’s groundbreaking Efficient
Coding hypothesis, put forward
in 1961 to explain how
the brain works at the
neuronal level.
The brain, Bar-
low hypothesized,
tries to form an
efficient repre-
sentation of the
environment by
switching on as
few neurons as
possible, thus
minimizing the
energy involved.
Prof Shi and his
team realized this
hypothesis was in-
complete, as it assumed
a passive organism, where
GROWING UP
the properties of neurons adapt to the statistics of
sensory input. What was missing from Efficient
Coding was that organisms actively shape and
optimize such statistics through their behavior. The
researchers added the effect of active behavior to
the hypothesis in work first presented at the 2012
IEEE International Conference on Development and
Learning and Epigenetic Robotics.
The team’s AEC framework takes into account
how animals and humans utilize their motor system
to facilitate the efficient encoding of relevant sensory
signals. An example is the simultaneous movement
of both eyes in opposite directions to align the
images from the left and right eyes so that they
can be fused into a coherent percept, also known
as a “vergence eye movement”. The framework is
a powerful unifying principle for the development
of neurally inspired and driven robots. It underpins
technology that will enable robots to become
more adaptive, structure their own behavior
automatically, and predict interventions
that match biological systems they
mimic. Prof Shi’s team has
already shown this single
principle can account
for the emergence
of a wide range of
other behaviors,
such as visual
tracking and
accommodation
(focusing), and
the automatic
combination of
multiple sensory
cues.
Prof Shi
e x p l a i n e d :
“In developmental
robotics, ideally we
want to put a robot in
the environment and let
The Active Efficient Coding
(AEC) framework utilizes the
perception-action cycle model
to advance development of
next-generation robots.
PERCEPTION-ACTION
CYCLE
Sensory
representation
in brain
Sensory input
Action
Like a baby,
a robot learns
to predict the
consequences of
its movement
and adapt to the
environment
through repeated
failure and
trying again.
PERCEPTION
BEHAVIOR
