Overlaps Between Biological And Artificial Neural Networks? 8
snurfer asks: "I'm very interested in the overlap between biological and neural networks, but I don't know much about the research being done in these areas. Where do people think the next breakthrough in neural network-type computing is coming from? Is there any possibility that improved understanding of neural function and how the biological brain solves problems and learns/remembers might flow into the next generation of software? I guess just in general, are the two research areas still borrowing from each other, or have they diverged too far to be of mutual benefit?"
The facts (Score:3)
Artificial neural nets are really great in a very small number of applications. They can particularly show good results for pattern recognition and can help you identify important criteria to solve a problem when you really have no other ideas on how to solve it.
But the field of AI research all but abandoned neural nets in the 1980s. They tend not to scale up well, they are simply useless for a large number of AI problems, and studying them tells you nothing about the human brain.
The little bit of information I can provide... (Score:3)
Re:The facts (Score:1)
This is not correct!
Neural Nets are alive. Have you considered the fault-correlations models, of computer networks
I read your chronology!
AI and Biology: Protien Processors (Score:1)
Four years ago i read an article( i am not sure but i think it was in "Scientific World") about the "Protien Processors". I forget the name of the protien but its property to react to laser-light brought it into the research labs. Contrary to binary state regime( 0 or 1) of semi-conductors this protien had the capability of maintaining multiple states: a vast potential for storing images/patterns and related image/pattern recognition etc etc. among other applications.
Russians got their first. They made a optical processor, based on this substance, some 25-30 yrs ago. They actually used it in Radars.
I read your chronology!
Re:The facts (Score:1)
I think one area which has caught the attention of researchers is the Network Management. I am at the Uni Of Stuttgart, Germany. Its traditional strenght has been telecomm systems. And i assure you there is a lot happenin here in regard to application of NN.
I read your chronology!
Re:The facts (Score:1)
Re:The facts (Score:2)
However, neural nets were at one point considered almost the 'holy grail' of artificial intelligence. We now know that neural nets are not and likely never will be. They are very good at solving a small range of problems but are totally unsuitable for many many problems.
Furthermore, the tuning that has been done to neural net algorithms generally makes them less like real human neurons, rather than more like them.
Many undergrads learn about neural nets and get all excited about the possibilities. After only a while of studying, this excitement typically ends. There is very little research still being done on neural nets (compared to the late 70s and early 80s) at the post-graduate level.
Artificial/Biological (Score:1)
Now, theoretically they are capable of the same things. By that I mean, traditional artificial NN's are capable of universal computation, just like any old Turing machine, which means that anything you can do on a regular computer you can do with an artificial NN and vice-versa. And biological NN's, one would suppose, are likewise capable of universal computation.
Theoretically.
In actuality, there are many things that you cannot practically do on an NN, because the equivalenet algorithm would run in exponential time, or training time would be enormous. So most things you see a regular computer do, you cannot do in an artificial NN due to exponential time requirements.
Conversely, you can do pattern recognition and a few other types of tasks incredibly easily with an artificial NN that are incredibly complicated to program traditionally. How would you write a computer program to classify sonar targets? A rule-based system would have thousands upon thousands of intricate and inter-related rules and would be very brittle, requireing thousands upon thousands of man-hours in programming effort. You can slap together an NN to do sonar classification (of limited types of targets) with maybe a week's worth of effort.
As to the relation b/n artificial and biological NN's: both consist of many neurons linked together. And in both neurons take all the inputs from all the other neurons in the network, weight them by some amount, add the values together, and then spit out some response. But there the similarities end.
In an ANN outputs are scalar values usually between -1 and 1. Each neuron takes as input the outputs of every other neuron in the network, weights them by a scalar, adds the weighted values together, and then puts them through a transition function to produce an output between -1 or 1. "Real" neurons (well, so it was thought) spit outputs of 0 OR 1 (no in-between) but backpropagation (the most popular training algorithm) requires a continuous output, and -1/1 outputs are mathematically convenient, so that is what is used. In most networks, the weights between many neurons are 0 (i.e., the neurons aren't actually connected).
Most ANN's use backpropagation as the training algorithm of choice, which implies the network is a perceptron, which means that the neuron connections form a DAG (directed acyclic graph). The lack of cycles limits the types of computation the NN can perform but makes training them possible (cyclic NN's are I believe still a wide-open research area).
Biological NN's work just like ANN's "in theory." Each neuron takes in a scalar input (an input voltage level), performs an operation, and spits out a scalar (an output voltage). However, it turns out that real neurons spit out voltages in "pulses" (so they spit out a voltage pulse w/ time-varying frequency). And it also turns out that neurons encode information in the frequency of the pulses, the time in between pulses, and who knows what else. So information content emitted by a biological neuron is much much more than simply the voltage level measured at a particular point in time.
What this means in practice is that biological neurons don't act _at all_ like artificial neurons. Artificial neurons read information ONLY from the level of voltage on the line at a point in time. Biological neurons are able to encode information in the time-varying aspects of their emitted voltage pulses. This means that in practice a biological NN is, possibly, capable of drastically higher levels of computation, per neuron, than an artificial NN. But as far as I know no one has any idea how this process works in a useful manner.
What this means is that you could not train or design a biological NN as you would an artificial NN and vice versa. Perhaps as more research continues we will discover more things about biological NNs and come up with artificial equivalents. But as it stands right now any artificial NN that is capable of being designed to perform a computation that its designers want it to perform does not resemble, except in the most superficial of ways, a biological NN.