DISQUS

kurt9 · 5 months ago

This is a step in the right direction. However, it is a long way from being able to simulate the complete functionality of neurons and dendritic connections.

GoatGuy · 5 months ago

I do wonder sometimes ... whether we research "dendritic connections" of neurons (as for an example), out of frustration with the lack of progress in the more traditional logic-and-coded-information based AI approach.

Animalia brains are made from obscurely functioned neurons, that store, that react, that live, that grow, that diminish, that die. Eyes are made of layer-upon-layer of neurons and chromophores that similarly work, live, die, transform themselves over a lifetime. Do we try to make foveal sensors for our digital cameras utilizing neuronal structures? No. We've come by way of logic to use square arrays of utterly fantastic (though now so keen as to be mundane by route of common ignorance) transistors, photosensors and digitizers. While not yet as power-efficient or resolute as the eye, fast we are coming to the point where they are.

Same holds for all these neurons-in-synthesis.

They're SIMULABLE, folks. Yes, perhaps the new memristors are simpler, more power efficient, whatever. But are the similulated zillions of them (there are simulated zillions, are there not?) showing that the reseach into neuron-synthetic methods is yielding brilliant results far in advance of anything logic-and-data based AI has already achieved? Are the rates-of-advance between the two communities of AI synthesis showing that the "neuronal" method is far more quickly increasing on an exponential or quadratic curve that is bound, in some reasonable period of time, to overtake traditional AI? Is it?

Because that's the metric. One cannot put money (or love, or much else) into the research of the ferro-magnetic memory idea (not really the memristor, but just saying), because simply put, the rate of increase in its density, power, and cost ... isn't on a curve to overtake conventional CMOS or NAND memory any time soon. Maybe never.

So, in like fashion, one needs to see the rationally predicted future of advance of the neuronal group - assuming eventual micro-device modelling replete with dendrites and all the rest - compared to the ever advancing L+D AI methods. Then, and only then, will I get all tingly regarding this stuff.

GoatGuy

nova77 · 5 months ago

@GoatGuy: Modern AI has long moved away from logic-and-coded-information. Even the example of the camera you make does not fit, since the face recognition uses a well known machine learning algorithm.

The question here is whether we will get strong AI via almost-perfect simulation of our brain or with statistical models such as belief networks (or equivalents).

GoatGuy · 5 months ago

OK, I'll buy that - "kind of". The methods used for face recognition range from Bayesian filtering to Markov chains, to subspace z-plane transform (followed by Baar learning matrices and so on. Insofar as I can tell, none of these were modelled after the way our own neurons are thought to process visual information. Similar, maybe ... but certainly I've not yet seen a paper that propounds that it has done the heavy lifting by way of modelling the preprocesing of the retina, the 2d integration of the visual cortex, and the 3-d autocorrelation of the medula. So far, not.

I remain of the opinion that the "simulate-a-neuron" (or a zillion of them) either by unique synthetic analog neurons (such as these memristors and memcaps) is moving along substantially faster than conventional computer science, or not.

GoatGuy

pater_leo · 5 months ago

It is cool of course but there must be made a next step to the biology point of view.
It is to implement the Viable System Model (VSM) of Stafford Beer - http://www.ototsky.mgn.ru/it/beer_vsm.html .
Suppose the time has come for this ! - http://www.ototsky.mgn.ru/it/presentations/VSM2...