|
INSPIRED
ELECTRONICS:
an interview with Rahul
Sarpeshkar
(this is an expanded version of the interview appearing in the print
copy of RLE at MIT)
2002
December Issue 1
RLE: What does the term "biologically
inspired electronics" mean?
Sarpeshkar: "Biologically
inspired" electronics refers to electronics whose design is
inspired by architectures or organizing principles seen in biological
systems. For example, a "silicon cochlea" design is inspired
by the traveling-wave architecture of the human inner ear or cochlea,
a "Reichardt motion sensor" is inspired by correlation
motion circuits in insect eyes, and a "biosonar front end"
is inspired by signal processing in the bat or dolphin. My research
inspiration comes primarily from neurobiological systems.
 RLE:
What are the features of neurobiological systems that indicate
possibilities for fundamentally new generations of devices?
Sarpeshkar: Neurobiological
systems perform complex sensory and sensorimotor tasks in real time,
with very low power, and in a very small volume. For example, the
human inner ear does at least a GFLOP of computation in a volume
that is less than 40 mm^3, and with a power consumption that would
allow operation on a AA battery for 15 years. Such incredible specifications
are attained because of a clever use of technology for computation
(fluid mechanics, micromechanics, and microelectronics in the case
of the ear), and through efficient nonlinear, adaptive, and distributed
architectures. Designs that insightfully mimic biological structures
hold great promise for ultra low power electronics, for feedback
system design, and for revolutionary computational architectures.
RLE: Why does hybrid computing
hold potential for dramatic power efficiency improvements in signal
processors?
Sarpeshkar: Hybrid computing
attempts to combine the analog advantages of low power operation
and superior exploitation of the technology with the digital advantages
of signal restoration, divide-and-conquer processing, scalability,
and programmability. Since such a computational paradigm combines
the best of the analog and digital worlds, dramatic power-efficiency
improvements may be possible
RLE: Why are such efficiency improvements
important for radically new types of electronic devices and systems?
Sarpeshkar: Power efficiency
improvements are important in portable systems, which need to compute
in a small lightweight volume on modest amounts of battery power.
Cell phones, laptop computers, medical implants in the human body,
wireless sensors, spacecraft, and mobile robots are all examples
of such devices. Many people believe that the explosive demand for
such devices will continue into the future and spawn applications
that we have not yet imagined. Power efficiency is also extremely
important in the scalability of future microprocessors to smaller
channel lengths
RLE: What does your groupís
recent success exploiting the hybrid computing approach to develop
integrated circuits that mimic neurological processes of the brain
suggest for the future?
Sarpeshkar: I believe that
the brain is the worldís greatest hybrid computer. It uses
ìspikesî or pulses to compute, which have an inherently
hybrid nature: the time between pulses is a continuous analog variable
while the pulses themselves are all-or-none discrete events. My
lab is researching the use of such time-based signal representations
to perform ultra-low-power analog-to-digital conversion, to build
analog memories, to build hybrid computers specialized for processing
sensory data, and to build novel event-based hybrid control architectures.
RLE: What are the most difficult
problems today facing scientists in attempting to recreate the capabilities
of biological systems in man-made devices and systems?
Sarpeshkar: The most difficult
problems lie in three areas. First, understanding how to perform
efficient-and-reliable computations with noisy and unreliable physical
devices, a feat that neurobiological systems perform routinely without
treating all computing devices as switches. Second, understanding
the design, performance, and robustness of distributed control architectures
that operate at multiple temporal and spatial scales. Third, advancing
technology to a point where we may begin to replicate the high fan
in and fan out capabilities of low-power, slow-and-parallel architectures
such as the brain, or to the point where we can combine various
technologies on one integrated substrate as in the human ear.
|
