I have written a neural network library. At the moment, it works fine, but I am not
sure how to format inputs and output in order to get the best efficiency.

The final use is at the moment an OCR.

This means, input are float numbers (various results of various transformations done
on the pitcure), and output shall be a letter.

For this message, I will reduce the input to two floats casted into 7 bits integers.

Shall I present:
- each 7b integer on 7 inputs, one per bit ?
- each 7b integer on 128 bits with only one bit set through out 128 ? (let say lsb
right, msb left)
- each 7b integer on 128 bits with all bits set from the least significant until the
one representing the current value ? (what means, for 0x14 I would set bits 0 to 20
starting from the right, what is 20 bits set up amongst 128)

For an 8b output, shall I use:
- strict 8 parallel bit output ?
- 256 b , taking care of the one with highest value ?
- how many bits from the right are set up ?

I do not care about efficiency at the moment, since my computer can do with 2 hidden
layers and 256 cells per layer fast enough for my application.

I just want to know what is the best for the network itself to learn well, and
recognise as good as possible.

Thanks for tips or URLs.

--
DEMAINE Benoit-Pierre (aka DoubleHP ) http://www.demaine.info/
\_o< If computing were an exact science, IT engineers would not have work >o_/

generic "neural network library" would have to output floating point
numberS (-S!), not letter (which is OCR-specific output)

apart from that, only recommendation that could possibly make sense, is
to organize inputs in a way so that different letters would result into
maximally different inputs. it all depends on your whole model,
starting from image transformations, and ending with NN type.