This is a very dicey question. Logic functions can be thought of as mapping multiple inputs to a single output. Now each logic function create its own boundary. So if you are using a complex logical equation it is actually very hard to approximate the underlying function. Here I am treating the input Booleans as the input features.
From practical experience: I had a n-bit = 8-bit input, and it mapped to a single bit output. I used a 2 layer Neural Net and then added a final layer of single node (to have a 0 or 1 output). I was using the sigmoid activation function along with normal back-propagation learning.
Now, I varied the hidden layer nodes from
1024. The cost decreased, indicative of correct learning but the accuracy did not change, howsoever I tried (changing learning rates, using momentum, etc). It was either giving all
1's or all
0's, even though I was using the same set for training and validation. The reason I hypothesized for this behavior was:
- The irrelevant input features are affecting the final output. Even after extensive training and testing on the same set with a huge number of hidden nodes.
- Due to the
0 nature of the input the NN was not able to create a correct boundary. Since if it created an equivalent feature of
x^3 using its nodes, it'll just output the same constant value again and again
f(1), where it should have been a nice little curve
So in short it might or might not work, you never can tell!