The issues of harmony have notably low profile in the current AI research. Meanwhile, when it comes to brain architecture, harmony plays the central role. As we hope to demonstrate in this essay, neural cells act primarily as local harmony detectors. On a less local scale, the consensus is that perceptions correspond to certain stable periodic patterns of neuron firing. A harmony-based architecture would probably be more successful than current AI paradigms.
It would help here to recall the basics of neuron firing theory. In the rest state neuron membrane typically has electrochemical polarization potential of 70 millivolts. When the firing impulse comes to the neuron from another neuron via the corresponding synapse (the site of their connection), this polarization potential changes, typically by 1-2 millivolts or less. If the polarization potential decreases beyond the threshold of approximately 60 or 55 millivolts, the neuron fires, otherwise the polarization potential tends to rapidly relax to the original rest level of 70 millivolts.
Hence, when the reception of an impulse via a synapse decreases the membrane polarization potential of the receiving neuron, we call this synaptic connection excitatory, because the decrease of the polarization potential makes it easier to fire for our neuron. Otherwise, the synaptic connection is called inhibitory. Because the reception of an impulse changes the polarization potential by at most 2 millivolts and because the polarization potential tends to rapidly relax back to 70 millivolts, the neuron can fire only if it receives several (from 4 to more than a dozen) impulses via excitatory connections simultaneously or in a very quick succession.
Hence the neuron works as a detector of several excitatory impulses coming almost simultaneously. So we can say that the neuron detects the harmony between its incoming impulses.
Now we shall turn to learning mechanisms in the brain, and observe that the local learning (on the level of one neuron) is directed towards detecting this harmony even better. As we have noted, the reception of an impulse changes the polarization potential usually by 2 millivolts or less. The actual value of this change is usually called synaptic strength. This value is not constant, but changes with time. This ability of synaptic strength to change is the key mechanism of neural learning and is called synaptic plasticity.
The most typical rule of synaptic plasticity for excitatory connection works approximately as follows. If a neuron fires shortly after receiving an excitatory impulse (i.e. the excitatory impulse contributed to firing), then the synaptic strength of the corresponding connection increases. However, if an excitatory impulse along the connection received, but the firing did not happen soon after that, then the corresponding synaptic strength decreases.
In other words, if the incoming firing was in harmony with other incoming firings (in sufficient harmony to cause the firing of our neuron), then the weight of the corresponding connection increases (our neuron is going to "pay more attention" to the "advice to fire" received via this connection). But if the incoming firing is not in sufficient harmony, our neuron tends to "pay less attention" to the future "advices to fire" from this source.
So locally neural cells work as detectors of harmony between incoming impulses, and local learning is directed towards enhancing their function as such harmony detectors.
[1] The source of all data is "Principles of Neural Science", Kandel E.R. et al (eds.), McGraw-Hill, 2000.
[2] For a typical example of what is wrong with AI neural nets see the popular Neural Simulation Language.
The neuron output is expressed via firing rate. Hence it is totally unsurprising that all harmony effects are missing. In fact, while some neurons capable of temporal summation of received impulses come back to the resting potential relatively slowly (20-100 msec), a typical neuron capable mostly of spatial summation of received impulses comes back to the resting potential very quickly (3-4 msec) [1]. At the same time [3] indicates that conscious perception is characterized by the typical firing rate of 40 Hertz and no more than 70 Hertz. Hence the intervals between impulses are typically 25 msec and no less than 12 msec in conscious perception. In such situation firing rates do not matter all that much, but what matters is whether the signals are in phase. This suggests one of key reasons of failure of AI neural nets to compete with biological neural nets so far.
It is also unsurprising that such networks exhibit high interference --- it is virtually impossible to train them to solve several different problems at once. As long as it starts learning something new, it forgets the previous knowledge quickly (with some relative exceptions, like ART-based systems --- see Chapter 8 from the link above). After all, it is the non-interference, which is THE "holographic property", and we should expect the phase information to play the crucial role here.
The typical training schemes are also flawed --- they are directed to globally optimize the network for one specific problem. This might work to obtain some ad hoc engineering solutions, but that's not how the nature works, and from the viewpoint of understanding and simulation of real biological neural systems it is a complete dead end. Equally bad is the widespread distinction between the "training mode", when the system learns, and the "working mode", when it does not change.
Real biological systems learn as they work and always keep learning, they can absorb almost unlimited quantity of various memories and types of problems to solve with very little interference (unless you cram in too much in too short a period of time), and while they might have some specialized learning mechanisms for especially important survival tasks, overall their learning mechanisms seem to be quite general and not specialized for narrow types of tasks.
These learning mechanisms are probably also some sort of harmony mechanisms, on the higher level, than a single neuron, but below the specific tasks. If the hypothesis of Crick and Koch [3] is right, the next level is those coordinated firings of neurons with the typical frequency of 40 Hertz (this characteristic value can vary between 30 and 70 Hertz). During such processes the whole groups of neurons seem to fire together. Since the typical difference in firing times of connected neurons should be in milliseconds (2-4 milliseconds seems typical), the typical chain for a cyclic signal should be from 3 to 15 neurons.
It is not difficult to see, that local learning rules for single cells support the storage of such patterns very effectively (subject for another discourse). However, I would conjecture that there are probably some even higher-level harmony mechanisms between this level and task-learning. So I suggest that learning of the specific tasks is the side-effect of a hierarchy of harmonization mechanisms.
[3] Crick F., Koch.C. (1990) Towards a neurobiological theory of consciousness. Semin.Neurosci., 2, 263-275.
[4] See also references in Spectral methods in the theory of consciousness (added in July 2001).
Mishka --- August 2000
Copying of this and my other papers on the science of consciousness is allowed free of charge, provided that the texts and this notice are unaltered, and that no further restrictions on the subsequent free redistribution are imposed
Back to Mishka's home page