Don't ask, don't get

[andrewducker]

I emailed the guy behind the rat brain experiment I mentioned yesterday.  I was curious about the way that training could be done without some way of feeding back that the right or wrong outputs were being received.  I was most gratified to get a response:

The neural flight control that is being reported is very rudimentary. The in-vitro network of rat cortical neurons simply controls the pitch and roll of the aircraft to produce straight and level flight, the neural equivalent of an autopilot. This is accomplished using an effect reported by Eytan, D., Brenner, N., and Marom, S., Selective Adaptation in Networks of Cortical Neurons. Journal of Neuroscience, 2003. 23(28): p. 9349-9356 in which "high" frequency stimulations (once every second) was reported to depress the response of the network while "low" frequency stimulations resulted in an enhanced response. For our system we tied the network's response to the control surfaces, dedicating stimulations on one channel for pitch, and a second for roll control. Each channel is stimulated separately, and the response (PSTH) is recorded. Control movements are proportional to the current error from straight and level by mapping the error (0 to 180 degrees) to the interval 0 to 100 ms of the PSTH and integrating the difference in response before training, to the current or enhanced or depressed levels. The more error, the more the control surface is moved. The networks only gradually control the aircraft since the Marom effect requires over 15 minutes to develop. The two frequencies are then used to adjust these weights (i.e. number of spikes in the PSTH) to produce optimal flight. The neurons/network don't seek optimal flight in the classic sense. Instead, we adjust the weights (using high and low Freq. stims) in the network to produce that result.

It is a very simple system and our only interest in it is in terms of those changes within the network and the possibility to extend it to more of the network than just two or three different channels.

Or, in other words, the neural net itself isn't learning anything - they're using a peripheral effect of the network to return a very basic response, and no actual learning is occurring.

Still, it's connecting neurons to electrial wires, which is a step in the right direction, just not as big a step as it first appeared.  Another case of hype over substance.

Comments

[code-delphi.livejournal.com]

Or, in other words, the neural net itself isn't learning anything - they're using a peripheral effect of the network to return a very basic response, and no actual learning is occurring.

It did seem a little too good to be true!

Still, it's connecting neurons to electrial wires, which is a step in the right direction, just not as big a step as it first appeared. Another case of hype over substance.

Anything for an eye-catching headline; "... always good for a gasp and a giggle amongst the jaded masses." as Julian May put it in Perseus Spur.

If such an interface could be wired into a human brain, could our brains learn how to use it? Both process stimulous received, and act upon the outputs? Could the rest of the brain "discover" the new sense/limb... wiring it into the processing centers where it'd actually be useful would be the real challenge.

This led me to wonder how the initial neural pathways are laid down in the brain in the developing foetus and early development. There's have to be some basic mapping (presumably), though where that'd be meaningfully encoded is the usual mystery (the junk between the genes in our DNA?). Alternatively, maybe our brains bootstrap themselves: all the inputs (nose, ears, eyes, nerves) and outputs are in place, and everything else—all the connectivity from which our conscience arises—somehow just develops in response to stimuli.

Regardless, a fascinating development; perhaps they'll do something a little more revolutionary in their follow-up work...

[heron61.livejournal.com]

If such an interface could be wired into a human brain, could our brains learn how to use it? Both process stimulous received, and act upon the outputs? Could the rest of the brain "discover" the new sense/limb... wiring it into the processing centers where it'd actually be useful would be the real challenge.

We have the answer to that and it is yes. Paralyze people can learn to move a cursor on a screen when connected to a implant to allow them to do so, and monkeys can learn to move robot arms the same way. Brains are incredibly adaptable and can rapidly learn to use new inputs, much like if you put glasses on someone that reverse everything, in less than a week, everything looks normal again because their optical processing centers alter their processing so that the images or similar reversed.