/ ai

Initial Impressions on Neuroplasticity and AI

Last month I read some really great stuff on neuroplasticity; Sam Kean's excellent 'Tale of the Dueling Neurosurgeons' right after finishing Norman Doidge's 'The Brain that Changes Itself'. So, I thought I'd take a bit of time to introduce you to neuroplasticity and perhaps attempt to point out places where knowing about it may tie back to artificial intelligence.

For the uninitiated, neuroplasticity is, according to the Oracle of Wikipedia:

"an umbrella term which encompasses both synaptic plasticity and non-synaptic plasticity -- it refers to changes in neural pathways and synapses which are due to changes in behaviour, environment and neural processes, as well as changes resulting from bodily injury".

Paraphrased, this means that it refers to the ways your brain changes in response to external forces, this is in contrast to the older view that once you become an adult your brain structure is pretty much set which is to say you become locked into your intelligence and way of thinking, and doing things. To be blithe, neuroscience without plasticity is the embodiment of being unable to teach an old dog new tricks.

Neuroplasticity argues that with training and time, nothing in your brain is ever set in stone. Both books use the catch phrase 'neurons that fire together wire together' and those that fire apart wire apart, so by training your neurons to activate together they will develop closer connections and smoother pathways so that you can use your brain in different ways, learn new skills and unlearn negative habits. One of the major researchers into neuroplasticity, Paul Bach-y-Rita became interested after his brother George taught their father to walk again after a major stroke by starting him crawling and gradually moving his way up, though an autopsy when his father finally did die showed the stroke had done major damage to his white matter communication cables the brain had reshaped itself to work around the damaged tissue.

I really found the biological results surrounding neuroplasticity extremely interesting, and so I did some cursory research to see if AI had found ways to incorperate it yet, and didn't find too much (admittedly, all this really proved was that I need to take the time to invest in a JPASS or something). On the flip side there is some literature you can take from artificial intelligence that isn't specifically related to neuroplasticity and relate back to the biological side. Those doing research and development on artificial brains like Blue Brain and Spaun could definitely gain some value from incorporating neuroplastic changes to their systems in the long run if their systems, by virtue of being brain simulations, don't already allow for change.

The central tenet of neuroplasticity already exists in artificial intelligence in terms of things firing together wiring together. Machine learning in general bombards a system with a lot of data and from there patterns often enforced are learned and those that aren't are not. However, I don't think it's frequently done (correct me if I'm wrong, my machine learning class seems so far in the past right now) that the entire structure of how the program learns is changed by what it learns. Especially since, in the human animal sometimes neuroplasticity, or plastic learning can lead to negative structures being built in the brain, such as gambling addictions.

We already have adaptive robots, I read a recent news article on Antoine Cully's team at Sorbonne University. They have built a robot designed to adapt to a broken limb. In Cully's system the robot has a number of preprogrammed sample gaits and evaluates which would be ideal to use to move given the limb is broken. Searching a finite set of gaits isn't a matter of a brain restructuring itself in robotics, but the resulting learned adaptive movements is one of the main accomplishments of the human and animal neuroplasticity and could perhaps become an eventual method to solve these problems in robotics. Humans would build these gaits through trial and error, and ones that worked would be reinforced in the brain.

The thing about neuroplastic change that really interests me is how divergent it makes brain structures out to be. It implies that there is not necessarily a 'right' way to make a brain, everyone's seems to function a little bit differently. In a way, it implies that you couldn't have an artificial general intelligence that didn't have some measure of 'unique personality' due to how its neural connections were formed.