The Magazine of IEEE-Eta Kappa Nu July 2017 - 22

FEATURE

Fig. 8. (a) Orthodox von Neumann computer architecture vs. (b) neural net. From Wikipedia.

a network of biological neurons. The circles represent
neurons, and the arrows represent the synapses
interconnecting the neurons. A neural network is
particularly good at matching patterns, based on a set
of training data. This strengthens certain synapses and
weakens others, creating a network that can learn by
example, without having to be specifically programmed.
This is an example of a neuromorphic or "brain-inspired"
computer architecture.
The neural net is not remotely similar to the von
Neumann architecture. Where are the logic, program
control, and memory? Both the logic and memory are
distributed among the synapses. While the von Neumann
architecture is a universal computer that can simulate any
other computer, even including a neural network, such a
simulation is not guaranteed to be either fast or energyefficient. Recent developments in artificial intelligence,
such as IBM Watson, have simulated neural nets with
many "hidden layers" of neurons between the input and
output, and have demonstrated "deep learning" using
supercomputers. This "cognitive computing" has shown
impressive results, but maybe there is a better way to
implement it.
Biological neurons are slow, noisy, and unreliable. The
characteristic switching time is of order ms, over a
million times slower than modern transistors. The power
consumption of the human brain is of order 10 W, which
is a million times less than that of a supercomputer.
Despite this, the human brain can be much faster than
a supercomputer at certain tasks related to matching
and recognizing patterns. For this reason, major research
efforts have been made into developing chips that can
emulate one or more aspects of brain structure. For
example, the True North chip of IBM and the Zeroth chip
of Qualcomm use conventional transistor technology,
22

where many transistors are needed to simulate a single
neuron. Other research efforts have focused on novel
device technologies such as memristors or spintronic
devices, which might emulate neurons more efficiently.
The field of neuromorphic computing is making major
gains [8], and will lead to a class of special-purpose
processors chips that are fast and energy-efficient for
certain types of problems. However, this does not mean
that we are actually designing or simulating biological
brains. We still don't quite know how brains work, and
ongoing research is showing complex structure on
multiple levels. But biology will continue to provide
inspiration to engineers in the future.

VI. Conclusions
As the classical approach of simply increasing transistor
count seems to be ending, the door opens to a wide
range of new opportunities, from new devices to new
architectures, from new applications to new types of
artificial intelligence. A central theme will need to be
energy efficiency. Several new approaches, including
neuromorphic and superconducting computing, are
discussed in this article, but other novel approaches are
also being actively developed.
The IEEE will be at the center of this endeavor, and
we at the IEEE Rebooting Computing Initiative are
encouraging both students and senior engineers to
participate - we need your enthusiasm and new ideas.
Participating ventures include the new industry-oriented
International Roadmap for Devices and Systems (IRDS),
and the student-oriented Low-Power Image Recognition
Challenge (LPIRC). The exact configurations of future
computing systems remain to be determined, but
continued growth in computer performance and
productivity is inevitable. The next twenty years will be
www.hkn.org //

THE BRIDGE

http://www.hkn.org

Table of Contents for the Digital Edition of The Magazine of IEEE-Eta Kappa Nu July 2017