Beauty is a Rare Thing
by
Chris Homan

It is remarkable that a formalism as powerful as a neural network so closely resembles the structure of the human brain. We would seem to be pushing our luck to expect it to function like the brain as well. The logical network, for instance, is also a simple and expressive formalism whose structure matches the brain's, but it is too unwieldy to function likewise. A functional brain model must be able to both represent large amounts of information in a small amount of space and organize itself in ways not directly related to its instinctual goals.

Neural networks satisfy both criteria. One important reason the brain cannot be a logical network is that it's synapse count is too low for there to be a one-to-one matching between brain cells and objects to be represented - there are no "grandmother" cells in the brain. The brain must represent information subsymbolically, that is, one brain cell must be part of the representation of many different objects. Neural networks, by decomposing information into canonical elements that can be recombined to represent many different things, economize information in the same way the brain must.

But the most striking thing about neural networks is that, like the brain, they can anticipate rewards before they actually occur, and use these expectations to make decisions. Reward-driven behavior is, after all the basis of human intelligence.

In case this is not apparent, consider that the underpinning of Freudian psychology is that all behavior is motivated by sexual desire. Although many of Freud's theories have long been suspect of possessing any therapeutic value, the fact that they form the basis of a popular genre of literary analysis proves that there is a significant degree of relevance in them.

Consider also that the human brain must have evolved to better enable propagation of the species. Evolutionary theory suggests that there is a physiological mechanism to reinforce decisions that ensure species success.

Data from neuroscience backs this claim. It has been observed that the body releases large amounts of dopamine after eating, mating, or fight-or-flight actions, actions that directly influence the ability of a species to survive. It is Dopamine causes synapses to become more conductive, and thus more influencial in future decision-making. In sense, dopamine has the effect of rewarding the brain for proper behavior.

But what about actions that do not directly affect survival? It appears the brain releases small amounts of dopamine when such decisions are made, indicating that it assumes the action will ultimately lead to a much bigger dopamine payoff. In this way "secondary" or "tertiary" goals can become more important than long term goals.

While neural networks are not the only computational systems to exhibit goal-oriented behavior - dynamic programming also does so - they are perhaps the only ones that can reorganize themselves in anticipation of a reward as effectively as the brain can.

To do so requires five main elements: a neural network, which takes as input a parameterization of the system's world, or state (this corresponds to the amount of excitation of each sensory nerve at a specific point in time) and outputs the reward at each state (corresponding to the amount of dopamine produced in the brain), a means of smoothing the reward values between two contiguous states by changing synaptic weights (which is analogous to dopamine's ability to effect brain synapses), and a means of indicating which states are success states (corresponding to the body's ability to produce dopamine to reward instinctual actions).

The neural net system decides which state to move to by picking from the set of possible next moves the one which has the highest associated reward. The synapses are then adjusted to minimize the difference in reward between the old and new states. What is most inimitable about this is that the changes happen at a subsymbolic level, just as in the brain, so that a promotion or demotion in the value of a state is instantly propagated through the system, in much the same way as a successful brain action becomes more influential than less successful ones.

 Another important characteristic of this kind of system is that it can be trained to perform tasks without needing to rely on hard coded rewards, in as much as the high-reward states are determined by the trainer of the system. This also mirrors human development. It can easily be argued that human behavior is determined as much by conditioning as by instinct.

 Again, a gleaning of several other fields of study seems to confirm this. Human development literature is full of debates on where nature, or instinct, ends and nurture, or training, begins.

Most striking, however, is the concept of structuralism - found in psychology, anthropology and philosophy - which states in essence that meaning is entirely dependent on cultural context, or the "training" an individual receives from birth.  While neither of these observations proves anything, they both confirm that the brain is a remarkably flexible organ, one that only neural networks have come close to emulating.

Neural networks show great promise as models of brain function because they exhibit the same kinds of space conservation and learning processes as the brain does. While more convincing evidence - for instance, a neural system which can "train" by example, the way humans can - would be needed to draw more solid conclusions, their rare combination of form and function make them a particularly exciting development.