Leaky ReLU

Plain ReLU outputs zero for every negative input, and its gradient there is zero too. A unit that drifts negative across the whole training set stops receiving gradient entirely and never recovers — a dead neuron.

Leaky ReLU leaves a small slope on the negative side so something always flows back:

1f(x) = x        if x > 0
2f(x) = alpha*x  otherwise

Task: write leaky_relu(values, alpha) applying this element-wise, rounded to 4 decimal places.

• alpha is typically small, around 0.01, but your code shouldn't assume any particular value.
• Exactly zero takes the second branch — alpha * 0 is 0 either way.

The leak is a slope, not an offset. Negative inputs stay negative and stay ordered, so the unit keeps some information about how negative its input was rather than collapsing it all to a single zero.