How can an engineer learn from neuroscientists
Mitchell Gil Maltenfort
mitchm at netzone.com
Tue Mar 4 01:36:22 EST 1997
In article <5fg1e2$bq6$1 at news-s01.ny.us.ibm.net>, junior1 at ibm.net [Bernie
>The neuroscientist view is not clear to me. I need some help here.
That's why I'm barging in again...
>I feel that some of the questions asked by a neuroscientist could be like:
>Why do we need to bother with engineers and simulate neural/sensory
>functions in silicon if we can work with the real thing?
As an engineer who works with neurophysiologists, I'd like to pick at this
question a bit.
The reason neurophysiologists are interested in models is that we often
can *not* work with the real thing...for one reason or another, the
experiments are not feasible to perform. For example, my own work was
looking at the ensemble behavior of a few hundred neurons - in the
physiological case, it's only possible to record from a few neurons
simultaneously. (Composite signals like the EEG are indirect
Think of physics, where theoretical analysis became more prominent as the
particles became harder to get at.
The question the neurophysiologists will ask are better phrased "Can we
represent the known behavior of each component of the system, and the
connections between components, so that we can simulate the system and
observe its behavior? Does the simulation then produce the important
behaviors expected from the physiological system, or do we need to develop
a better understanding of the structure so we can improve the model?"
For a neuron, the components might be the ion channels, the cell membrane,
the dendritic tree. In a reflex, the components might be the sensors
(say, pain receptors), the afferent fibers carrying information to
interneurons or to neurons innervating the muscle, the interneurons and
the motor neurons.
I say 'might be' because the relevant details might change depending on
the question the scientist wants to ask and the researcher's own opinion
of what features are important. Can a population of neurons be lumped
into a single mathematical function? Can firing rate be represented as a
continuous-valued function, or is the timing of neural firings important?
Do we need to represent the dendritic tree in detail or can we use an
equivalent cylinder description?
Neurophysiology is like any other engineering problem in this respect: you
have to be sure you define the constraints of the solution appropriately.
More information about the Neur-sci