Neural Networks

General Cedric

Has anybody here used a neural network to control a robot? I was thinking about using a neural netowrk programmed in python.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
-General Cedric


"Wikipedia is your Friend"-Me

kwinn

What hardware (CPU) are you going to use? Neural nets of any size require a lot of processing power, and using python (or any other interpreted language) would multiply that requirement.

Leon

That is true for the training phase. Once the network is trained, though, it can run on much less powerful hardware.

Leon

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Amateur radio callsign: G1HSM
Suzuki SV1000S motorcycle

soshimo

What are you using for an activation function in your hidden layer? If you are using sigmoid or some other non-linear function then you will definitely be slamming the processor. Remember, you don't have a dedicated math coprocessor so all those floating point calculations are going to take memory and/or cycles. If you go with a step function then you are limited to only a perceptron type NN which usually has binary classification (no fuzziness). I've tried to get a NN to run on a PIC uController which has a lot more RAM than the Prop and pretty much you would have to dedicate a very high-end chip just for the NN itself. At least those were my observations based on RAM and the processing power of the chip itself (which was 33MIPS).

LilDi

Take a look at the CogniMem. great products, but pricey

http://www.roadnarrowsrobotics.com/

soshimo

At a 110 bucks a pop you don't want to make any mistakes [noparse]:)[/noparse], but they are nice. I was reading the specs, 10us recognition time max (for a 1024 matrix) and they have direct digital video recognition. Can you say true vision? If I ever win the lottery or my boss decides to get a wild hair, maybe...

Here's a link directly to the chip