Ben, On Wed, Aug 29, 2012 at 12:31 PM, Ben Goertzel <[email protected]> wrote:
> > Picking one particular tiny illustrative detail of this - my realization >> that neurons MUST communicate derivatives like dP/dt rather than straight >> probabilities, to be capable of temporal learning without horrendous >> workarounds. I thoroughly explained it on this forum, and no one objected >> to any of it, yet it has changed nothing. >> > > To those of us not working on neural net models, this sort of insight is > kinda irrelevant... > I'm not so sure. Don't you use Bayesian methods to compute probabilities, that change with circumstances? If so, by converting the inputs to dP/dt notation, computation throughout remains in dP/dt. At the outputs, you will need to integrate to get back to P. The net result is that you no longer need past memory for temporal learning - all without a single neuron-equivalent in your code. > > But still, this is an interesting observation. > > It reminds me of work studying neural population coding using Fisher > information > > http://prl.aps.org/abstract/PRL/v97/i9/e098102 > > [Fisher information being an average of the second derivative of a > probability density, it's kinda like the derivatives you reference...] > > I'm curious: How would you modify, for instance, the Izhikevich neuron > equations > I wouldn't. This is SO simple, it is almost hard to even see it. Converting to/from dP/dt is something you do at the inputs and outputs. Within the NN/Bayesian "network" computations remain the SAME, only now things become naturally capable of temporal learning. Of course, with ANY such "simple" change you MUST lose SOMETHING, and here is is: No longer will the network learn based on merely the simultaneous presence (or absence) of things, where the inputs don't change (nearly) simultaneously. For most things, this is a MAJOR improvement, but keep your eyes open for this "loss". > > http://www.izhikevich.org/publications/spikes.htm > > in accordance with your idea? (I reference this just because it's the > neuron model I've worked with most recently.) > All the evidence I have seen is that spiking neurons are an "implementation detail" to facilitate carrying information over long axons. The spikes themselves are relatively uninteresting beyond the information contained in their separations. Some bandwidth is lost to conversion to spikes, but not as much as would otherwise be lost in the long axons. There are MANY short non-spiking neurons in our brains, not counting the glial cells. The spiking neurons are the ones with the long axons. If you ever get a chance to look closely at an exposed brain, you can see what looks like tiny glistening fibers. These are bundles of long axons. > > Regarding your idea for a cross-disciplinary math/AI/neuro research > institute -- I wish I had the power to get something like that formed. > NO ONE has this power. THAT is why it will take a group to push this through. The trick would be to get 2 or 3 leaders from each of the 3 disciplines, and train everyone to tag-team discussions with potential hosting countries. The idea would NOT be to usurp anything in a hosting country, but rather to support their efforts as part of the overall project. It might be easiest to simultaneously sell this to both the technology and the health people, as new cures will doubtless come from such an effort. The presence of the scanning UV fluorescence microscope (or other similarly capable technology) would greatly facilitate this. Right now, the health people have more money. Maybe I'll be able to do it in a few years time, in HK or China or > Singapore, we'll see... > I wasn't really expecting a return message to report of China 8-:D> It would probably take a year or so just to recruit the people and get a robust plan together, to even start to "shop it". I suggest that you start looking for the people who would be the nucleus of such a project. You should probably recruit about twice the number you will need, as many will fall by the wayside rather than putting in the work needed to forge a good plan. When you finally have a group and a plan, get everyone to start sending it to the agencies who might fund at least part of such a thing. Alternatively (the easier way for you) send people you think might be a useful part of such a project to me, and I'll coordinate the forging of a plan for prospective funding. [Hey you nerds, wanna go work on a REALLY BIG project far away?] Note that like fresh fish, THIS IS PERISHABLE. In a few years, unless my sense of history is WAY off (don't bet on it), AGI will become every bit as much of an anathema as perceptrons became. Note that "perceptron" isn't even in my spell checker. Creating such a group will become MORE difficult with time, because funding agencies will no longer see the pot of AGI gold at the end of this rainbow, so there will be less money available. IMHO the one BIG plus in AGI is that it has the potential to attract money - if only the feasibility issues can be addressed. The envisioned project would seem to satisfy at least some of the feasibility concerns. From a funding POV, AGI and the envisioned project are MUCH stronger together than they could ever be separately. Steve ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
