Signal Processing - November 2016 - 10

Figure 2. Arindam Basu, an assistant professor at NTU Singapore's School of Electrical and
Electronic Engineering, holds a low-power chip that connects wirelessly to neural implants.
(Image courtesy of NTU Singapore.)

Wireless neural implant
connectivity
When embedded in a brain, neural
implants can be used to alleviate the
debilitating symptoms of Parkinson's
disease, give paraplegics the ability to
move their prosthetic limbs, or alert
epileptics to an impending seizure.
Researchers at Singapore's Nanyang
Technological University (NTU)
recently developed a chip that connects
wirelessly to neural implants, enabling
more efficient, accurate transmission of
brain signals (Figure 2).
For prosthetic patients, a neural
implant is typically connected via
wires to a computer that decodes the
brain signals and enables the artificial limb to move. Yet external wires
are not only cumbersome but require
permanent openings into the brain,
increasing the risk of infections. The
new chip is designed to resolve both
of these problems. Project researcher
Arindam Basu, an assistant professor at NTU's School of Electrical and
Electronic Engineering, says that tests
on data recorded from animal subjects
showed that the chip could decode the
brain's signal to the hand and fingers
with 95% accuracy.
Raw neural data can easily amount
to multiple megabits per second.
10

Transmitting this level of data reliably will drain an implant's battery
within a matter of hours. Designed
to be highly power efficient, the chip
analyzes and decodes data at the
implant site before compressing the
results and sending it wirelessly to
a small external receiver. "We can
reduce the data rate to tens of kilobits
per second, increasing the battery life
a lot," Basu says.
"The chip consumes ten to 50 times
less energy than off-the shelf processors," Basu says. "We achieved this
[capability] by adopting a mixed analog-digital processing approach as
opposed to the fully digital processing
used in commercial processors," he
notes. Most of the processing is performed in analog to reduce energy
while a small, but critical, fraction of
computing is accomplished in digital to
achieve accuracy. "The signal processing algorithm was also chosen carefully
to suit this approach-a randomized
neural network, like extreme learning
machines, was selected where a large
bulk of the weights or coefficients are
random and fixed, which can be easily
implemented in analog without loss of
accuracy," Basu says.
The chip is also designed to analyze
data patterns and detect any abnorIEEE Signal Processing Magazine

November 2016

mal or unusual patterns. "A pattern
recognizer is quite general and can be
trained to detect any pattern in a signal that may have diagnostic value,"
Basu says. "For example, for epileptic
patients with medically incurable epilepsy, this chip can be trained to detect
the electrical onset of epilepsy before
the clinical onset kicks in." The capability can be used to trigger electrical
stimulators to suppress the seizure and
can also notify a caregiver about an
impending seizure. The chip can also
be used to track arrhythmia patterns in
electrocardiograms, which are indicative of heart abnormalities.
The researchers opted for a two-layer
neural network based on "random projection to higher dimension" in a first
hidden layer followed by a trainable second layer. "As long as input dimension
D is much larger than the number of
classes C, the number of multiplications
in the first layer (DxL) is much larger
than that in the second layer (CxL)
where L is the dimension of the hidden
layer," Basu says.
"The key observation is that the random number multiplications in the first
layer can be implemented in very low
energy using analog circuit as opposed
to traditionally used digital approaches
(~10-100 times lower)," Basu explains.
Analog circuits, however, suffer from
non-idealities such as mismatches or
statistical variations. "This is not a problem in the first stage where the weights
are supposed to be random," Basu
notes. "In fact, we exploit the mismatch
to create the randomness." However, to
achieve high accuracies, the researchers kept the second stage processing
in the digital domain, enabling it to be
separately trained for the particular
mismatch distribution of every chip.
"Thus by making a careful choice of
algorithm and careful partitioning of
analog and digital computing, we are
able to achieve high accuracy as well as
low energy," Basu says. "Project development and testing processes are ongoing," Basu continues.

Sweat monitor
Many medical professionals view wearable sensors as an affordable and highly

Table of Contents for the Digital Edition of Signal Processing - November 2016