Hi, > Using artificial rules, such as hardball winner-take-all and > synaptic weight normalization, it's doable to get ANN's to do this. > > But in an autoassociative network with realistic biophysical > properties, controlling activity to prevent runaway synaptic > modification is a very large problem. > My own grad advisor, Mike > Hasselmo has worked on this very problem using pharmacological > modulation to suppress synaptic transmission during learning. > The fact that epilepsy usually starts within the hippocampus > (with its sheets of 100k neurons, all interconnected with > excitatory connections) indicates that this is a real problem for > the brain as well as models of it.
This is analogous to problems we've seen in Novamente's probabilistic inference engine, actually. The analogue of "runaway synaptic modification" in Novamente's inference module is "runaway truth value modification", in which the system arrives at a set of links representing probabilistic relationships in reality, and then keeps iterating on these and revising them (revising the premises based on the conclusions, then revising the conclusions based on the premises, etc.) until the accuracy of the model degrades. Fairly careful adaptive dynamical control is needed to keep these problems under control. The fact that similar problems occur in Novamente inference as well as in the brain, suggests that they're "general system-theoretic problems" in some sense, perhaps occurring in any distributed network-oriented computing system. > They keep the real data too, but it's *huge* (100+ channels of > 70khz data, realtime). The raw data is basically an average of > the neural activity of the nearby cells. Spikes from neurons > within a small radius of the electrode tip stand out and have a > certain characteristic shape/amplitude, which is used to identify > said cell. Apart from identifying spikes, I'm not sure you'd get > much out of the raw data(assuming you are also collecting EEG > data realtime at 10khz or so from one electrode in the nearby region). > > However, nowadays people are starting to worry about complex > spikes too (bursts of spikes). Assigning these spikes to their > source neuron is much harder because spikes after the first one > in a burst are reduced in amplitude. So you need specialized > clustering algorithms that are aware of bursts and what they do > to a spike amplitude. You need to go back to the raw data to > identify such bursts every time you change your detection algorithm. Interesting. As you probably know, "spiking neurons" are the latest fad in the formal NN community. So the computationalists are trying to keep up with the neuroscience results... e.g. http://research.microsoft.com/~cmbishop/pulsed.htm is a good book (though not the newest stuff) -- Ben ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
