ambati@acf5.NYU.EDU (FJLevM{n[]Balamurali Ambati) (02/01/91)
Is it correct to say that people have had more success in modelling biological phenomena (i.e., in this case, describing neuronal networks in the visual pathway / hippocampus / ...) than in designing networks to solve problems that the visual pathway / hippocampus / ... can solve? A simple example is the immense difficulty in making a computer "see." Of course, one of the problems is that "seeing" involves much more than the visual pathway alone. But is this the only problem? It's my understanding that Hopfield-Tank and other similar neural network models are not that useful (when compared to existing digital algorithms and even some genetic algorithms) in obtaining near-optimal solutions to combinatorial optimization problems such as TSP, etc. Is this because these models are simplistic in terms of describing the appropriate neurons? Or is this because the human brain was not designed to solve TSP, etc.? Is it worthwhile making neural networks that can themselves invent specific algorithms (somewhat like humans make machines solve specific problems)? Is it possible / simple? Balamurali K. Ambati
ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards) (02/01/91)
In article <1467@acf5.NYU.EDU> ambati@acf5.UUCP (FJLevM{n[]Balamurali Ambati) writes: >Is it correct to say that people have had more success in modelling >biological phenomena (i.e., in this case, describing neuronal networks >in the visual pathway / hippocampus / ...) than in designing networks >to solve problems that the visual pathway / hippocampus / ... can >solve? That's a loaded question. For the most part, we don't know a great deal of hard information concerning what is really going on in (let's say) visual processing in humans. We think we know alot about the retina, but past that we don't have the hard facts. There are alot of high-level description theories (e.g. Marr's) of what is going on, but there is controversy even over high-level descriptions. The problem is multi-faceted: 1) There are so many neurons involved in these processes 2) They are packed in together incredibly 3) Fan-in and fan-out can be in the range of 1000-10,000 for each neuron 4) Not every brain is the same, and there is some evidence that small scale brain structures can chagne slightly over time 5) We are just barely getting a complete picture of what a single neuron does, but there are around 1000 different neuron species You see, we can't pick out neuron by neuron of a human brain and figure out what is going on computationally. We have to seek and understanding of brain at a level of neural organization instead of neuron connections. In other words, we have to answer how can populations of neurons organize themselves (operationally or genetically) to produce information processing structures. >It's my understanding that Hopfield-Tank and other similar neural network >models are not that useful (when compared to existing digital algorithms >and even some genetic algorithms) in obtaining near-optimal solutions >to combinatorial optimization problems such as TSP, etc. Is this because >these models are simplistic in terms of describing the appropriate >neurons? Neural Networks are usually used to form non-linear mappings between and input space and an output space. This is a very general problem to solve, so it is reasonable that they do not perform as well as specially crafted TSP programs. I don't think there is anything you can say psychologically about TSP results from neural networks. Why? Well, we don't know that brain is organized in any manner like the neural net models which we put forward. However, there is also no reason that traditional neural net models can't resemble neural population organization, but there is no positive evidence that they do. But neural nets are still capable of learning some neat things for themselves, but they need to be a bit more specialized before we see anything like proto-intelligence comming out of them. Some of the most interesting "neuromorphic" work has come from looking at a part of the brain which we do have alot of understanding of... the retina. We know that lateral inhibition is used in the retina to provide an increase in dynamic range of illumination across the retina. It could also be used to bring out edges. Retina-model image preprocessors have shown themselves to be of great use, especially with things like IR detectors which have a tendency to differ in response from detector element to detector element. There is also evidence that brain uses spatial frequency information for visual processing. Of course, engineers have been using Fourier transforms for a long time to do spatial-frequency domain processing. But more importantly, I think we have to understand the non-linear dynamics of neural circuitry. Feed-forward neural nets are wonderful for to problems, but the real world has a temporal component which we have to deal with. Also we have to integrate heuristic learning mechanisms from traditional AI which work so well on "symbolic" processes in with traditional neural net models which handle non-well-behaved information well. -Thomas Edwards
gowj@novavax.UUCP (James Gow) (02/04/91)
Are there any references to work on graphical representations in kb's? linc james