|
| The neuron-based computer: Andrew Dou, University of Illinois Urbana-Champaign |
Tens of thousands of living brain cells have been used to build a simple
computer that can recognise patterns of light and electricity. It could
eventually be used in robotics.
A computer built using tens of thousands of living brain cells can recognise
simple patterns of light and electricity. It could eventually be
incorporated into a robot that also uses living muscle tissues.
Artificially intelligent algorithms inspired by the brain called neural
networks have been used for everything from chatbots to searching for new
laws of physics. Normally, these algorithms run on conventional computers,
but Andrew Dou at the University of Illinois Urbana-Champaign and his
colleagues wondered whether they could instead use actual living brain cells
– neurons – as part of the set-up.
The team began by growing around 80,000 neurons derived from reprogrammed
mouse stem cells in a dish. The process was similar to that used for
creating brain organoids, also known as mini-brains, which are clumps of
neurons that have been used as simple information processors, as well as for
studying intelligence itself. The main difference is that the neurons in the
new device were arranged in a flat, two-dimensional layer.
To complete the computer, the researchers placed the neurons below an
optical fibre and onto a grid of electrodes so that the neurons could be
stimulated with both electricity and light. The electrodes could also detect
when the neurons produced their own electrical signals in response. All of
this was housed in a palm-sized box, which, in turn, was placed in an
incubator to keep the cells alive.
Conventional neural networks can easily learn how to distinguish different
patterns of signals, so the researchers attempted to train the living
computer to do the same.
They first created 10 distinct patterns of electrical impulses and light
flashes. To train the computer to recognise them, they played the 10
different patterns repeatedly over the course of an hour and used a regular
computer chip to record and process the electrical signals that the neurons
produced in response.
The neurons produced the same signals each time the same pattern was
presented. The chip, which was running an artificial neural network, just
had to learn to distinguish those signals.
Often artificial neural networks can take a long time and many iterations to
train, but the division of labour between the neurons and the chip, a method
called reservoir computing, allowed the researchers to minimise this.
Overall, the whole procedure took less time and energy this way, says Dou.
After the hour of training was up, the researchers let the neurons rest for
30 minutes, then exposed them to each of the 10 sequences of light and
electricity again.
To evaluate how well the device did, they calculated a performance score
called F1 that is commonly used for neural networks, where 0 is the worst
possible score and 1 indicates perfect pattern recognition. The device’s
best score was 0.98.
Dou says that in early experiments it could not score above 0.6 because the
neurons would sometimes produce electricity unexpectedly, due to naturally
occurring random processes. To improve it, he and the team worked out a
combination of chemicals and additional electric impulses to suppress the
randomness.
The new device is an early step in the researchers’ long-term goal to
develop living computers and robots. In the past, they have made robots that
use
muscle tissues to move, and other researchers have used living brain cells to make robots
process information
about and
navigate through their environment.
These experiments used a computer to relay signasl from a robot to the brain
cells, which were not in direct contact. However, incorporating neurons into
a robot would mean that the neurons could sense their environment, then
process those inputs at once, with less mediation, says
Nicolas Rouleau
at Wilfrid Laurier University in Canada.
“Brain cells sort of micromanage themselves. They connect to each other on
their own, and then you can give them information in many ways,” he says.
Neurons respond to pressure, chemicals and magnetic fields in addition to
light and electricity, so they could take in lots of information about a
robot’s environment all in one go, he says.
Using living cells for computing, and reservoir computing in particular,
makes for energy-efficient devices that can also keep working even if some
of their smaller parts get damaged or experience failure, says
Ilya Shmulevich
at the Institute for Systems Biology in Seattle. Consequently, a robot that
mixes living neurons and reservoir computing could have advantages over more
conventional, purely mechanical machines, he says.
At the moment, the device cannot compete with conventional neural networks,
but Dou says that the team is going to make it larger and more complex in
the hope that some unexpected behaviour, or behaviour that they didn’t input
into or train the network for, will emerge as more and more cells interact
with each other.
Dou
presented the work at the American Physical Society’s March Meeting
in Las Vegas, Nevada.