# How to use WISARD neural network to detect defects in banknotes

We learned about the WISARD neural network in my machine learning course. We said that for an $n\times k$ image, we would use discriminators having $n$ RAM neurons each of $2^k$ bits. The examples we covered were always simplistic and involved small images, something like $3\times 4$.

I am looking at some past papers in preparation for my exam, and in one particular question we are asked to describe the architecture we would use for a WISARD neural network to detect defective bank notes, having a resolution of $768\times 384$. I'm thinking that $768$ RAM neurons of $2^{384}$ bits each is not feasible. What architecture should be used in this case? Are we perhaps expected to group pixels together to have less neurons with less bits?

Any help is greatly appreciated.

• Note that the number of atoms in the universe is about $10^{80}\approx 2^{266}$ so your intuition that $2^{384}$-bit neurons are infeasible is quite an understatement. :-) – David Richerby Jun 12 '18 at 14:19
• @DavidRicherby Exactly, our lecturer frequently reminds us how many atoms there are in the universe, so I'm sure he's not expecting us to use $2^{384}$ bit neurons :) – sigma Jun 12 '18 at 14:21