Page 160 - MODES of EXPLANATION

Basic HTML Version

Two Case Studies
Understanding without explanation in neuro networks
The first lab I want to consider was trying to understand the mechanisms of learning at the
network level: the network properties of living neurons. This was a fundamentally new
research paradigm when it was introduced in the early 2000s, when studies of learning were
customarily done on single neuron interventions. The principal investigator (PI) of this lab
felt that because the brain is a network of neurons, to get a grip on an understanding of
neuron behavior there was a need to construct dishes of living neurons. The PI developed an
eight by eight grid, called the multi-electrode array, onto which embryonic wrap neurons are
plated. They are first disassociated and then allowed to grow their own connections. The
resulting dish is a generic model of basic neurological processes in the brain. This living
neural network is then embodied by connecting it to some kind of robotic device. The
overarching research aim is to develop the control structure for goal-directive learning in this
embodied network.
The sexiest embodiment of the dish is called MEart, a mechanical drawing arm that is
effectively a robotic artist that draws through feedback loops. It lives in Australia and
communicates with the United States via satellite. MEart has also traveled around the world
and exhibited its art in Paris, St. Petersburg, and China. MEart began with freehand,
spontaneous drawing as a response to neural activity in the feedback loop. The research
question was: “Can MEart learn to draw within a box?” If it could, then the research team
would have controlled its behavior.
Three researchers were involved in the project: D2, D4, and D11. D2 was the one
doing the robotics and he traveled the world with MEart and other embodiments. D4 was
conducting open-loop electrical stimulation, trying to get the dish to do what they wanted.
D11 worked on solving a very significant problem: getting the dish to be quiet, as it was
bursting with spontaneous network electrical phenomena.
Early in the research process there was very little interaction between the physical
model – the dish – and the computational models, as each researcher was working on his or
her individual goal. However, after the computational model was sufficiently developed, the
researchers began to interact with one another quite actively, and this was when they started
developing the mathematical representation they needed in order to control the dish. The
spontaneous network electrical phenomena across the dish looked random, like noise. The
bursting prevented the detection of any systematic change due to controlled stimulation; the
signal was drowned out by the noise. Consequently, there was no way to detect learning in
the dish.
So the researchers formed a hypothesis: learning required the bursts to be quietened.
They hypothesized that they could do this by providing a substitute for the natural sensory
input, via electrical stimulation. After over a year, D4 was finally successful in getting the
dish to be quiet. The bursting stopped – but then nothing happened. Every time they tried to
give the dish stimulation so that it could learn, it would drift into another pattern of
spontaneous bursting. They called this the
problem of drift
. It was such a significant problem
3