Some simple models of neurons

linear neurons

\[y = b + \sum_i x_i w_i\]

Here \(y =\) output, and \(x_i = i\)-th input. The sum is over the input connections.

binary threshold neurons

\[z = b + \sum_i x_i w_i\] \[y = \text{ if } z >= 0 \text{ then } 1 \text{ else } 0\]

Can rewrite the bias to be a threshold in the if statement (threshold should be -b).

rectified linear neurons

\[z = b + \sum_i x_i w_i\] \[y = \text{ if } z >= 0 \text{ then } z \text{ else } 0\]

sigmoid neurons

\[z = b + \sum_i x_i w_i\] \[y = 1 / (1 + \exp(-z)) = S(z)\]

where S is the sigmoid function.

Simple example of learning

Learning handwritten digits by learning templates; winner is the template that has the most overlap (weighted) with the written digit.