Hey, here's a post that I've been thinking about for a few weeks but decided to write while still half-baked due to the analog/hybrid computing thread that's been going on.
om One of the recent trends in computational neural science is to take the spiking behavior of neurons into account, just search "spiking neural network". Okay, so how does spiking enhance neural computation? Traditional neural networks are purely spatial in nature. That is, each neural ensamble is fed a vector of values representing its stimulus. In a spiking neural network, time is introduced. This might seem to have an obvious meaning, but the real value is slightly more obscure. One of the major problems in computer vision is that a neural ensamble trained on a specific stimulus must be presented a candidate match for that stimulus in a very similar position and orientation, otherwise it will fail. A spiking neural network can solve this problem by applying a space-time rotation to the stimulus. That is, the spatial pattern is transformed into a temporal chain of spikes. The recogniser for this chain of spikes does not care WHEN the chain begins to arrive, only that when it arrives, that it have the appropriate structure. A re-positioning of the stimulus can thus be ignored because it only causes a time-shift in the signal but no change to its structure. Another interesting feature of this system is that the brain can begin responding to stimuli the instant they happen. The stimulus can be thought of as a signal to which the brain can react at any time. It might be tempting to think "Aha! Spike train good! Me make spike train!" =P Well, there you go. That is pretty much a logical fallacy, that of confusing the mechanism with the function. It is sufficient to merely understand what the function is, how you go about implementing it is completely up to you. In general, you should tend to select mechanisms that you feel comfortable with. om Now with regards to computation. Neuromorphic hardwarez is all the rage these days. Indeed, in the past week there have been a bunch of articles that I noted but didn't read touting how so omfg brain-like the new memristor (r-ram)-based neuromorphic machines are going to be. I'm not even sure why this is news this week, just goes to show how little I care. =P Now, why don't I care? I actually have a good reason for not caring. Sure, it will be nice and low power and analog and you are basically simulating a synapse with a 7nm grain of olivine crystal. But, u know what? It reeks to me like premature optimization! [insert entire body of literature on digital vs analog here] summary: fast enough digital can do anything any physically realizable analog can do but with all the advantages of digital; in future some selected subroutines that demand the fastest possible latency might be analog while the Real Work will be done digitally. om Okay, so what then is consciousness? Consciousness can be thought of as a signal placed in a delay loop. In the brain we have the most important super-structure of them all, the cotico-thalamo-cortical loop, that passes through the striatum and basal ganglia on the way. The Thalamus is most famous for injecting information into the CTC loop, the cortex runs pattern recognizers against it, then the striatum synchronizes and filters it, where it then goes back to the thalumus and completes the circle. I would claim that the design-space for systems that will exhibit consciousness are fairly broad, each with varying characteristics. The main goal of the system is to maintain a "tagged world model" that is error checked against the perception stream... GAH! I've done gone spilled the beans there, with that much inf0z, even u knuckleheads should be able to build an AI. =\ -- IQ is a measure of how stupid you feel. Powers are not rights. ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
