Sigmoid
Things to note:
\[y = \frac{1}{1+e^{-x}}\]
- Max is 1, occurs around \(x=4\). Min is zero, occurs around \(x=-4\). Not a great range. When \(x=0\), \(y=0.5\). Sigmoid works like a probability since it returns a number between 0 and 1.
Tanh vs. sigmoid
Things to note
- Max is 1 but min is -1. At \(z=0\), \(y=0\). Grows faster than the conventional sigmoid. Note, sigmoid is just a shape. Technically tanh is also a sigmoid function, but data scientists mean the first equation when they say sigmoid.
Linear activation
- Not really an activation. It’s what it sounds like. \(y=w*x + b\). Might show up on the exam.
ReLU
\[Max(0,x)\]
Things to note:
- No saturation problem like sigmoids, but it can have the issue of a dying-relu, meaning that when the learning rate is too high, or when a neuron learns a large negative bias term for its weights, the ReLu will be stuck yielding zero all the time since it only takes negative inputs. That neural becomes is thus useless.
Leaky ReLU
\[Max(.1x, x)\]
Things to Note:
- There is also a version called parametric relu, which is \(max(\alpha x,x)\), where \(\alpha\) is a learnable parameter. I think this is the same \(\alpha\) as the slope in gradient descent, but I’m not positive.
