Feature map — bias and ReLU

Background

After the convolution sum, a conv layer adds a bias and applies a nonlinearity (usually ReLU) to the whole feature map:

$$\text{out} = \max(0,\; \text{feature\_map} + b)$$

ReLU keeps the "pattern detected" (positive) signal and zeroes the rest, so depth doesn't collapse into a single linear map.

Problem statement

Implement feature_map_relu(feature_map, bias) that adds bias to every entry and applies ReLU.

Input

• feature_map — array-like 2D of pre-activation convolution outputs.
• bias — float.

Output

Returns an np.ndarray the same shape: max(0, feature_map + bias).

Examples

Example 1 — the lesson's grid (bias 0)

Input:  feature_map = [[-2, 3], [4, -1]], bias = 0
Output: [[0 3], [4 0]]

Explanation: negatives clip to 0.

Example 2 — a negative bias raises the threshold

Input:  feature_map = [[1, 1]], bias = -2
Output: [[0 0]]

Explanation: $1 + (-2) = -1 \to 0$.

Constraints

• Add the (scalar) bias to every entry, then max(0, ·) elementwise.
• Return an np.ndarray of the same shape.

Notes

• The bias shifts the ReLU threshold: a more negative bias makes the detector pickier (only strong responses survive).