Image from: Rick Bolin |
It seems to me that the most exciting thing is to know what
is possible in the near future. I have been interested in artificial
intelligence for years, but I was still surprised to get to know that we are
now so near to an artificial brain.
IBM and HRL are making a neuromorphic microprocessor as
capable as a cat brain by 2016, which will be the outcome of the SyNAPSE project started in 2011.
This may seem trivial given that IBM Watson, a supercomputer
equipped with intelligent software, has already beat human in chess and the
quiz show Jeopardy. It is now being promoted to be applied in medical care as well. Nevertheless, the breakthrough of SyNAPSE is to have an intelligent
chip instead of a supercomputer. The technology of artificial neural network
has long been widely used, but at a huge cost for high performance. IBM Watson
is as large as several rooms, and your three meals each day cannot keep it working
for even one hour. Realizing neural network at the chip level does make a big
different here.
The basic idea is to use a fundamentally different
architecture when building the microprocessor. Traditionally all computers
follows the Von Neumann structure, where the system has separate units for
memory, processing, control and input-output units. Information is processed
sequentially. In a neuromorphic chip, however, memory, processing and control
units are mixed, and a high level of parallelism is realized. I have thought
that this kind of circuit would inevitably require new materials and device
characteristics, but I really have overlooked the potential of the current
transistor based integrated circuit technology.
IBM is building such processors by implementing neuromorphic
algorithms in the hardware design. Many of these algorithms have already been
used in software. The “hardware coding” makes the system much more efficient
than just running intelligent software in normal computers.
In an introduction video, Dr. John Arthur, one of the researcher from IBM, mentioned “current
computers can do a fantastic job in adding numbers, but they do really poorly
in recognizing faces where human brain is very good at.” The goal is to build
systems that can do recognition tasks automatically.
In the same video, Dr. Horst Simon, the deputy director of
Lawrence Berkeley National Lab made a very interesting analogy to the invention
of plane to describe this objective. Unlike birds, the planes have rigid wings
instead of flapping wings. Dr. Simon said:
“Reorganization computing is exactly at the stage where we are
looking at “flapping wings” and “rigid wings”. We don’t want to build a bird;
we want to build a device that allows humans to fly. So, we don’t want to build
here a human brain; we want to build devices that can help solve the tasks that
current computers cannot solve at ease.”
“We don’t want to build a bird; we want to build a device
that allows humans to fly.” I found this comment quite inspiring: take whatever
reference to help finish the task. I always wonder what Dr. Von Neumann’s
reference was when he outlined the architecture that gave birth to our current
computers. But now, our brain is the natural and powerful reference.
No comments:
Post a Comment