Unsolved Problems in AI (Part 6): Brain Plasticity

Our brains are no computers. Not even close.

Contrary to modern day computers (even the latest and most flexible generation of neuro-morphic computers) and claims of many AI researchers, our brains cannot seriously be understood as rigid "computing" devices, but have to be seen and studied as living organs composed of billions of living cells which constantly self-adapt - triggered by our experiences, the physical and chemical environment (internal and external) and even our own thoughts.

From birth to death, our brains are never the same on two different days of our lives. Recent estimates vary over the amount of change that happens in our brains, but it is quite safe to assume that about 60%-70% (!) of the brain cells of our neocortex change their connections every single day ! That is, many billions of nerve cells adapt every single day in the way how they connect to and interact with the other cells in the brain. This is an enormous and very powerful ability and not at all well understood by our artificial intelligence community so far.

This astonishing ability of our brains to self-adapt is called: plasticity.

By plasticity I refer here only to a few clearly identifiable brain features (there would be many more to discuss, but I focus here just on the main features of living brains that fall under the plasticity concept) such as:

1) synaptic plasticity: this is the nerve cells ability to change and adapt their own synaptic connections between pre-synaptic neurons and post-synaptic neurons. This is the most studied and best understood plasticity feature of the brain. It is the basis for the so called "Hebb Learning" model which states that neurons that "fire together, wire together". According to Hebb, neurons that show time correlated firing and excitation patterns, auto-adapt their synaptic connections so that the likelihood of future firing together is increased and strengthened. With this process the nerves "learn" to adapt and to react to internal and external stimuli.

This Hebbian learning rule and its various implementations in current neural network models (such as Artificial Neural Networks, Deep Learning, Convolutional Nets, Restricted Boltzmann Machines etc) is usually implemented by the standard back-propagation gradient descent algorithms. But these standard algorithmic implementations are just very simplified approximations (more like bad caricatures) of the real biological chemo-electrical processes that go on in the brain needed to achieve the powerful synaptic plasticity. In standard AI models the synaptic modifications are mostly modelled by numerical "weights" (i.e. usually real numbers between 0 und 1) allowing the modification and adjustment of the connection strength between artificial neurons during the "learning phase".

However, recent studies indicate that the exact sequence of excitation (not just the correlation) of the pre- and post synaptic activities is important for the modification of the synapses and also the specific time window in which they occur (around 40 milliseconds). Current AI neural network models do not reflect these important insights yet.

It is also known today that neurons do not only communicate through their adaptive synapses but also interact and communicate with so called glia cells that are important to establish the physical support environment for neurons. It is not clear yet if and how much glia cells are essential for cognitive processes.

In addition, the neurons use signalling spikes and so called "spike trains" to communicate which leads to a spike-timing plasticity which is barely understood so far and not implemented in any current standard AI model (as they usually need differentiable neuronal transfer functions, but the biological neurons use spikes and spike trains or bursts of spikes which are not differentiable functions when modelled truthfully).

2) developing the connectome: besides the just discussed synaptic plasticity another very important feature of the brain is its ability to not only modify the connections the nerve cells have already established among each other by Hebbian Learning, but also the ability to create such connections among cells and the ability of the cells to find other cells to connect with in the first place !

The neurons develop the so called connectome which can be compared to a neural highway network for signal transmissions among neurons that are not close to each other. Neurons do not only actively connect to other neurons in their immediate vicinity but often also to many neurons quite far away in completely different regions of the brain (on the scale of the size of neurons a distance of say10 centimetres in the brain amounts to a huge distance for the neurons as the brain cells are usually only a few nano meters big). These far reaching connections between different brain areas allow for higher functional cognitive abilities and even the conscious processing of the nerve signals.

3) flexibility: the ability of brain regions to take on or take over very different functionalities of other brain regions or expand or reduce brain regions dedicated to execute certain tasks when needed.

It is well known and studied today that for example the brain of a musician, say a violin player, changes with the amount of practice. The regions of the brain that represent and control the fingers of say the left hand of a violin player who practises much will grow larger than the same area in an average person's brain that does not play an instrument. This applies to many brain areas when their related functions are repeated and practised more often than usual.

Another example: professional long term taxi drivers have usually more detailed and larger brain areas used for navigation and recognition of visual scenes than average people have and it is well known that blind people make use of the unused visual brain areas to improve and aid their hearing and auditory processing. Some blind people can even do echo location similar to bats and identify objects merely by their sound reflections they generate with click sounds from their mouths. With lots of practise, some blind people can even hear and analyse the sound reflections of fully hidden objects inside of other closed objects like a suit case inside the closed trunk of a car. So, by using the unused visual processing areas of the brain to improve their auditory system, a blind person can detect more by not seeing than a "normal" person can see with both eyes open!

The amazing ability of the brain to change and deplore functionalities of certain brain areas indicates that the brain has a kind of homogenous underlying architecture so brain modules are able to take over functionalities from different other brain regions. This would seem impossible if the different brain regions had completely different underlying architectures. The processing regions of the brain seem to have a fractal architecture that is self-similar in most regions of the neocortex and other brain regions.

4) graceful degradation: is the brain's ability to continue working despite substantial damages or loss/damage of brain tissue.

For many thousands of years we have seen cases of people with badly damaged heads and brains, especially in war times. Brains can also be damaged massively in no war times by for example: accidents, tumors, infections, drugs, poison, strokes, birth defects, temporal suffocation etc. Over the last two centuries or so such brain damages have been scientifically studied in much detail - with some astonishing results.

Considering how essential our brain is for our overall functioning and the control and maintenance of our body functions (homeostasis, blood pressure, digestion, body temperature, hormones, movements of our body parts, perception and interaction with the world etc) one would expect that even the smallest damages to the brain would inevitably lead to death or at least to major lasting damages to our brains and bodies.

However, we are now aware of many documented cases of patients in which substantial parts of the brains were destroyed (and hence not functioning) but sometimes without barely any noticeable effects or symptoms ! People may have major brain damages, but as a medical layman one would not notice, unless one were explicitly told about the damages and their symptoms.

Even in cases where the brain damage is indeed obvious and noticeable with clear symptoms, for example after a stroke (one sided paralysis of the body, speech problems etc), the brain still often completely or at least partially recovers over time and the symptoms can disappear. This can happen even though the damaged brain cells are not functioning anymore as they have died off or lost their connections to other cells. This is only possible because other areas of the brain can take over some or all the tasks of the damaged brain tissue (see flexibility).

A classical example of graceful degradation is often seen when doctors cut through the so called corpus callosum. The corpus callosum is the main nerve fiber bundle connection between the two brain hemispheres. It contains around 200 - 300 million nerve fibres. These fibres are sometimes cut by neurosurgeons when patients have frequent and strong epileptic seizures. To stop the seizures it helps when the connection between the two hemispheres is cut so the seizures cannot easily involve both brain spheres. After such a surgery people have two disconnected brain hemispheres in their sculls, however surprisingly the effect on the behaviour of such patients is barely noticeable in day-to-day situations. Try this with any kind of computer: cut the main data connection between its processors and see what happens.

5) change stability: despite the above described various structural and functional changes in our brains during our lifetime and the negative effects of ageing and later age diseases, the overall functionality of the brain usually remains in tact and remains mostly stable.

This is a major achievement of the brain as new nerve cells are generated in massive amounts during different stages of brain development and age and have to be integrated in the existing nerve network without disturbing the nerve network functionality. During pregnancy a child generates on average several million new neurons every hour !

Its an unbelievable logistical performance of the brain to handle all these new neurons, put them at the right places in the brain and then even integrate them in the already existing neuronal network of billions of other cells. This massive self-organising feature of the brain is one of the most difficult problems to solve and understand in AI. If we ever understand how the brain does this and have an algorithmic model for this, we could then use such algorithms to solve any real world logistic problem with ease.

On the other side, nerve cells constantly die and sometimes also in massive amounts at certain developmental stages of the brain. This is called "neural pruning" and happens for example in childhood prior to puberty and also later prior to myelination when the neuronal axons are wrapped in fat to speed up the neural signalling within the neurons. Neural pruning is necessary to reduce energy consumption as it erases neurons that are no longer needed and hence keeps the energy use of the brain low.

We do not know yet, how all these brain functions can remain intact or even improve despite the changes in structure and complexity it undergoes during a lifetime.

Just for fun, compare all this with a typical computer where even any single bit error usually leads to a crash of the whole system !

6) Chaos and the brain: Despite its general stability just discussed, in another sense our brains are also chaotic systems and very fragile in their functioning. Even minor changes in its chemistry (prove: just drink a good glass of whisky) or structure can sometimes have tremendous effects. Traumatic experiences (for example witnessing a murder) can have dramatic consequences for our brains and may cause a cascade of physical, chemical and hormonal changes inside our brain without any physical or chemical cause from the outside. Sometimes even a simple but intense nightmare can have such dramatic effects. We need no bad physical experiences to cause major changes in our brains, sometimes just our own thoughts can cause these (this is also the reason why placebos work).

In healthy individuals dramatic changes in our brains are actually rare and when they occur the brain usually quickly adapts and "repairs itself" (as described above). It seems the chaotic system of our brain is creating strange attractor loops in which it stays and holds our conscious and circles around until disturbed and forced out of the strange loop just to create new and more strange attractors to temporarily settle in.

I have to leave it here with this. I hope, I made it clear that our brains can hardly be compared to a computer. It should also be clear that when AI researchers are referring to our brains as blue prints for standard AI models that this can only be understood as a bad joke or a caricature of the real thing.

Saying that current AI hardware or software models are modelled after our biological brains (a claim often made by the big industrial players in the field) just shows no understanding of how the brain actually works and is an insult to all our brains and our intelligence.

Much more efforts need to be spent to understand how our brains manage to work so well while they constantly change and adapt and how the most complex self-organising processes work that make this possible.

This in my eyes is a key task in trying to understand the emergence of intelligence.

E.Schoneburg

Hong Kong, Berlin

November 28, 2016