Basic Intelligence and the Amazing Ability of Neurons
Intelligence is a topic that has been discussed on this site recently. Defining and describing intelligence can be a difficult thing, there are no shortage of definitions. There are obviously levels of what we may call intelligence going all the way from bacterial behavior to consciousness and self awareness. I will write a series of articles on intelligence, starting at the level where organisms can learn, and call it basic intelligence. Asking whether simpler organisms or evolving systems of such organisms have basic intelligence is a valid point, however I won't go into that here.
I believe a lot can be learnt from studying and considering how neurons behave and adapt and will draw the readers attention to some surprising and very important results in this regard. This kind of approach makes sense as you can then see how intelligence developed without having to approach everything at once. There are simple organisms with basic intelligence that it makes sense to study without having to solve all the problems associated with consciousness, intent, attention etc. So in this article, I will talk about and attempt to define what I call basic intelligence. That is a level of intelligence that is useful, but without all the complication of things like the self. In later articles I will discuss higher levels of intelligence with concepts introduced here playing a central part.
Defining something can be difficult as "intelligence" is just a word, meaning something slightly different to each person. However in order to talk about basic intelligence it must be defined in such a way that it satisfies many peoples understanding of what it is, while still being consistent. Rather than launch straight into a definition, I will leave that until near the end, after the motivation for such a definition will be clear.
Firstly I will point out what I regard to be essential qualities of basic intelligence then some surprising examples to illustrate what is going on.
Qualities of basic intelligence
So what could be considered qualities for basic intelligence?
1. A system capable of performing calculations such as with neurons or a computer. We know enough about brains and intelligence to know that the ability to calculate is essential for a simple brain.
2. The ability to adapt to new situations with meaningful behavior.
3. Some kind of pattern recognition skill.
4. Nothing to do with self awareness etc.
The amazing abilities of neurons
Please watch the video in the following article that demonstrates such basic intelligence, and pay attention to the behavior of the neurons, not the engineers.
Briefly what this shows is neurons automatically adapting to a completely different environment to what they would normally have been exposed to. They are rat neurons, but they are in a jar, nowhere near a rats brain, and processing information that is input in a way nothing like what would happen in an actual rats brain. In spite of that they learn and adapt their behavior to do something useful. It is not exactly clear from the article what was needed to make this learning start, random electrical stimulation seems to have been all that it took. What was not needed was some high level instructions to tell the brain to avoid walls. There was nothing to tell the neurons what the signals meant, but they were still able to extract information from what was input, and generate meaningful behavior.
What you should realise is just how flexible and adaptable neurons are. We cannot yet get a computer to display such general learning behavior.
An even more striking example of such flexibility involving rewiring sensory input is described here and here.
"All reasonable doubt that the senses can be rewired was recently put to rest in one of the most amazing plasticity experiments of our time"
In the experiment discussed, the ferrets brain was rewired so that the visual sensory input from the optic nerve was hooked up to the brain regions where hearing is normally processed. In spite of this, that brain region automatically configured itself to make sense of the visual data, and even allowed the ferret to see with 20/60 vision (33% of the normal 20/20). The structures that normally are found in the visual region of the ferrets brain automatically formed in the section normally used for processing sound, and did so well enough that the ferret was able to see. How was this achieved?
The design of the experiment allows you to rule out many possibilities. Genes in the brain, or specific brain chemistry can be completely ruled out because the auditory section would have the wrong genes expressed and brain chemistry. Some kind of top down instruction can also be ruled out because the ferrets brain doesn't "know" that such rewiring has occurred. The only possibility is that the neurons were able to automatically recognise the pattern, and adjust their internal structure to match it and extract the relevant information. It also should be noted that unlike in the first rat brain experiment, at the beginning of the learning process there was no feedback whatsoever given to the neurons as to whether they were doing the "right" thing. There was no training run where they were told the right answer, all that was required was the pattern in the electrical signals coming in the optical nerve for them to latch onto and "figure out".
As of today, psychologists, computer scientists, mathemeticians and engineers do not know how neurons perform this feat. However what we do know is that the neurons did not use rational thought or mathematics as we know it to achieve this. There is not contained in each neuron all the mathematical formulae needed to figure out vision such as edge detection etc. For all we know such abilities may be mathematically intractable. The structures required for vision grew in a bottom up, self organizing manner.
A general algorithm for basic intelligence?
As was mentioned in the articles, there was a debate in the psychology literature up until about late last century as to whether brain regions were specialized, requiring specific genes, brain chemistry and needing to develop in the correct place, or whether the brain adapted itself to the input. Obviously genes and chemistry have a use and have some effect as the ferret in question had 20/60 not 20/20 vision but as far as the debate was concerned these results argued overwhelmingly for the brain being very flexible and adaptable.
Now this flexibility suggests that there is a general algorithm for basic intelligence that is shared by all neural tissue in sufficiently advanced animals. If that is indeed the case, once it is figured out how neurons form connections then if attempts to put this structure and function in another substrate such as silicon are successful then AI would make significant and rapid progress. Such a system would in the order of weeks or months be able to do make more progress in some AI problems than computer scientists have been able to achieve in decades.
A definition for basic intelligence
Based on the observations above, we can now give the following definition for basic intelligence that describes what it is from a signal processing and systems perspective:
"A system that automatically adapts to a signal it has not been trained to, and extracts useful information from it demonstrates basic intelligence. The more varied and complicated the signals it can adapt to, the more intelligence it demonstrates."
In the case of the ferret you can easily tell that useful information has been extracted because it could see, however there is another important way. The internal structure of the auditory section of the ferrets brain grew to match that of the visual section. The fact that this happened automatically demonstrated basic intelligence and showed that such information had been extracted. Now there will be some subjectivity about what is a more complicated signal and which classes of signals are more important to others, but obviously a system that just detected edges from visual data would be less intelligent by this definition than one that did that, and built up a 3-d model of the world as a result.
In order for this definition to be useful higher intelligence also should not be defined in a radically different way to this. It is reasonable to expect higher intelligence to be built out of a system that can demonstrate basic intelligence in such a manner. A beating heart is made up of heart cells that automatically beat in time, and it is much easier to build a strong structure out of strong basic materials than weak ones, therefore I expect higher intelligences to be constructed out of systems that can by themselves demonstrate basic intelligence as defined here.
How may this have evolved?
Here I offer ideas on how and why such an ability evolved. To start with, consider the roundworm C. elegans, which is well studied and has about 300 neurons. These neurons cannot adapt to input in the way a ferrets brain does and the connections are pretty much determined by its genes. The question is what caused the neuronal connections to be no longer explicitly determined by genetics?
This was a very important transition because it allowed the beginning of what I have defined to be basic intelligence, or what you may also call learning. The most obvious answer is that it happened for the same reason that the position of each individual air sac in your lungs is not determined by genetics. Even if it was physically possible, there simply is not enough room on the genome for it to be stored. Instead genes if expressed properly form the fractal branching pattern that ends up growing those air sacs. No doubt a similar thing happens for heart cells, liver cells etc. As the sensory organs increased in size and complexity more data needed to be processed, the number of neurons increased to the point where the wiring diagram could not be stored in the genes. It needed to be learnt based on the input data. However there may be another cause that came first.
Even with a wiring diagram completely specified, there would be errors which could be fatal to the organism. Any kind of error correction ability that evolved even before the sensory organs grew too big to be handled by about 300 neurons would have a definite survival benefit. Such error correction would allow neurons to automatically link up in the correct manner. With error correction in operation, parts of the genome used for wiring could be freed up for other uses. The amount of sensory data to be processed could then be handled just by increasing the number of neurons, without storing a wiring diagram rapidly growing in size.
When such an event occurred it was significant in the evolution of intelligence and arguably life itself. Whether or not you call the roundworm intelligent, it is clear that the neurons even in a small part of a ferrets brain show a level of intelligence clearly and significantly greater than that of a creature with fixed connections between its neurons.
Defining and talking about intelligence can be a difficult thing to do, however like many other things it can be studied at a more basic level, which makes discussing and defining it easier. The experiments involving rewiring sensory input into a part of the brain it is not normally connected to have profound implications for the understanding of how brains work, the very nature of intelligence itself and likely AI. The evolution of brains that can automatically learn from input data rather than having their connections hard wired was a significant step in the evolution of intelligence and life.
Many ideas in this post have been inspired by reading On Intelligence, which I will discuss further in later articles when talking about more advanced intelligence.