Robot + Neural Network Learning to Follow Light

[keithosmarferrer]

hahaha, this is very cool, btw, i'm making a robot that uses a neural network that i have created, note that i used analogue materials and no programming, i have written some schematics logic and neural maps, here is the website of my robot: letsmakerobots(.)com/node/3042­9 just remove the() from the dot

cheers! btw, i'm 14!

[rommel241]

Oh I know, and its very cool, anything that can actually showcase the abilities of NN's is more than I have at the moment :P.

[shaggyman87]

by known solution do you mean a network that already has the correct weights, or an algorithm that describes the correct behavior?

[pirobotproductions]

@shaggyman87 Yes, I mean a network that already has the correct weights. I feed the current light sensor inputs to that network and the outputs from that network become the target values for the untrained network.

[rommel241]

@pirobotproductions Doesn't that defeat the purpose of learning? Your not even correcting its inputs, your just saying 'be like him', and correcting the waits accordingly. No learning there at all, just slow progressive copying. Your end result has no further capabilities than whatever the teacher can do, no possibility of learning, for example, a noisier data set.

Why bother learning at all once you have the teacher? :P

[pirobotproductions]

@rommel241 Yes, you have a good point. This is a form of "guided learning", much like how a tennis or golf instructor might grab your arm and move it through a correct swing. The goal is then to have your brain map your arm's proprioceptive sensor readings into your own motor commands. So yes, you could call this copying rather than learning. Also, it was really meant just to be a simple demonstration. :-)

[koz4q]

@rommel241

Keep in mind that even teaching a NN this way will result in a totally different network with totally different weights and although it may act identically in many situations, it will probably act differently in others. You can take a bunch of such NNs and average their outputs to get a more accurate result.

[rommel241]

@koz4q I understand that, but given that the robot has probably not learned anything about the problem itself, the edge cases you describe will probably be handled in random ways by the bot, as it doesn't have any developed knowledge about the system, only what it has copied.

If it happens those random ways are positive, than that was just by chance, and not attributable to the NN or learning process at all.

[x89codered89x]

hot shit!