Focal Loss

The problem, from first principles

Focal loss addresses class imbalance by down-weighting easy examples and focusing on hard ones. It is particularly useful for object detection where background is dominant.

Where p_t is the model's probability of the ground truth class. When p_t is high (easy), the weight (1-p_t)^gamma is small. When p_t is low (hard), the weight is large.

From theory to code

Implement:

focal_loss(logits, labels, gamma=2.0)

Where logits are raw model outputs, labels are 0/1 class labels.

Constraints

• logits and labels same shape.
• gamma >= 0.
• Return scalar loss.