I'm really interested in understanding the intuition behind multilayer perceptrons and neural networks.

I'm following the Caltech video which is excellent https://www.youtube.com/watch?v=Ih5Mr93E-2c

More specifically it goes through an example at 30 mins in where a circle classifier is approximated by intersections of straight lines. More specifically consider the lines

$l1: y-x=1$, $l2: y-x=-1$, $l3: y+x=1$, and $l4: y+x=-1$

Now suppose that i split each space into 2 depending on whether a co-ordinate point is above the line or not call these regions $h_{i}$ and $\overline{h}_{i}$ respectively. Then i know i want to classify positive if i'm in the square intersection of all 4-lines and so if the variable

$\overline{h}_{1}h_{2}\overline{h}_{3}h_{4}$ evaluates to true. They didn't go through this in the lecture but how would i represent this in a neural network like they did at around 27 mins in for the XOR problem?

Any help would be appreciated i really want to try and get a feel for breaking these problems down.


Okay, It's been a long time since I tried to build neural network on paper, but I think it would look something like this: (the red lines have negative weight, the blue ones positive, the threshold-neurons take the inputs and output zero if the sum is 0 or less and 1 if it's positive) enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.