Contrastive Loss

#deep-learning #loss-in-ml #interview

This is the first loss used for Contrastive Learning
Only one pair of data points are given
- The points are similar or different

Contrastive Loss Formula

$L o s s = (1 - Y) * | x_{i} - x_{j} |^{2} + Y * m a x (0, m - | x_{i} - x_{j} |^{2})$

Here,

Y = 0 if sample is similar
Y = 1 if sample is different
m is a hyper-parameter

This loss mainly minimized the Euclidian Distance between similar points and increases the Euclidian Distance between different points till some limit m.

Related Notes

Discriminative vs. Generative Models
F-Beta Score
Multi Layer Perceptron
L1 vs. L2 Regression
Vanishing Gradient