From cs240@cs.rochester.edu Tue Nov 4 17:59:36 2003 Date: Tue, 4 Nov 2003 17:57:25 -0500 From: cs240 To: tetreaul@vienna.cs.rochester.edu David Ganzhorn: Assignment 3 David Ganzhorn
cs240: Assignment 3
dg003m@mail.rochester.edu
9/25/03

Neural Networks as a Possible Human Brain Model

Computable models of the brain tend to take one of two approaches: over simplified models that are not complex enough to produce very interesting results, or more accurate models that are so complex that only small portions of the brain can be modeled. These later models generally involve the modeling of the behaviors and interactions of neurons. Neural networks have had much success at solving certain small problems, but there no neural network has been made that even begins to approach the full complexity of a human brain. However, the success of neuron models so far, and the fantastic success of the neurons within a human brain, indicates that there is much potential for such computational structures. The knowledge we have about the way the brain functions is clearly far from complete, but I believe we know enough about neurons in order to say that a very good computational model can be constructed my imitating the structure of neurons in the brain. However, the inability to implement our limited knowledge of neurons into electronic systems prevents answering the question of whether or not a neural network can model the brain.

The structure of a neuron seems simple at a glance: input comes in through dendrites, output goes out through the axon. However, this view greatly simplifies the whole process of a neuron. The inner workings of a neuron are amazingly complex; a complete copy of the bodies genetic code is encapsulated within each neuron, and the behavior of the neuron is in part dictated by the ongoing process of activation of certain genes within that cell. Beyond the changing factors within the cell, there are numerous factors occurring outside of the cell as well. Some chemicals, when introduced to the brain, can have dramatic changes on the overall behavior of the system, especially certain drugs such as amphetamines or hallucinogens. Even more mundane factors such as a person's level of hydration, blood sugar level, or blood oxygen level can greatly impact the brain's functioning. In short, neurons are subject to a vast number of influences that can greatly affect their functioning. Although it is possible to ignore many such factors when creating neural networks, there may be certain critical factors in the brain environment that allow neurons to be the computational components of the brain. The many receptors on the surface of a neuron are constantly interacting with the cellular environment, and perhaps it is the combination of the firing of synapses with the influences of the cellular environment that allows for sufficient computational complexity to arise from the brain. This lack of understanding of the functioning of neurons makes it impossible to say whether or not a neural network can model the brain, but what we do know does not yet eliminate the possibility. Perhaps neural networks will just need to be many times more complex before they can reproduce the functioning of a whole brain..

A great limitation of neural networks is their limited size; brains have roughly 10^11 neurons, and 10^15 connections between neurons. The number of connections is far greater than even the most sophisticated computer chips available. This issue combined with the simplified models of neurons means that successful neural networks will likely need millions of times more nodes, and each node will have to be many times more complex. Because both difficulties compound each other, neural networks may need to achive billions of times more computational power.

The overall architecture of the brain poses another problem for neural network models of it; how does one create a model with billions of individual units that each are connected to thousands of other units, and not only can the strength of the connections change, but new connections to units can be made as well. Also, the chemical environment throughout the brain is something that I believe has not been studied thoroughly enough for an accurate model of the chemical concentrations in all subsystems of the brain to be established. Although the cellular environment may not play as large of a role as the network of billions of neurons, I believe it will still play a critical role that must be accounted for in order to create a successful model of the whole brain. Also, the brain is a three dimensional structure, where as most chips are primarily two dimensional; it will be a challenge, if not impossible, to implement proper connectivity among the nodes of a neural network powerful enough to model a brain, on a two dimensional plane.

One seeming advantage that computer chips have over biological neurons is computational speed; current processors run at billions of operations per second, and neurons cannot cause spikes more than about one hundred times per second; it would seem that you would be able to vastly simplify the overall structure of the brain by using fewer, much more powerful processors to model it. However, this idea has a critical flaw; the brain isn't a large serial processor. Nearly a hundred billion neurons are all doing something simultaneously, and are constantly interacting with each other, being influenced by and influencing other neurons continuously. Perhaps it is impossible to model such a system with even a million ultra-fast processors if they can only perform in a serial mode, as one hundred billion slow processors operating in parallel with each other is an unimaginably complex system. Further, the timing of the individual spikes from neurons is thought to encode more information on top of the frequency of spikes, so the communications between all of these billions of processors is also of undetermined complexity.

Although there are several great difficulties that are involved with creating a working model of the brain, I personally feel that if a sufficiently complex model were created, that it would produce equally if not even greater intelligent behavior. We know a great deal more about the brain than we can yet accurately model with electronics, as we don't have any computer chips with hundreds of trillions of transistors, let alone such chips specifically designed to work for neural networks. It can only be expected that as computer technology advances, neural networks will be able to perform in ever more successful ways. I think it is clear that it is impossible at this point to say either that neural networks are either capable or incapable of modeling the brain, but that it is certain they will come closer to this goal in the future. Because of the several compounding difficulty factors in modeling the brain, I believe it will be a relatively long time, several decades at least, before any method is found to model the human brain.