Siamese network and triplet loss

Siamese network is an architecture which runs two networks with shared weights (effectively runs the same network twice) on two different inputs simultaneously. It is commonly trained with a contrastive loss such as triplet loss in order to draw together the representations of similar inputs and push apart the representations of contrasting inputs.

Define distance as the norm between the two encodings:

$$d(x_i, x_j) = \lVert f(x_i) - f(x_j) \rVert^2$$

Goal: learn parameters so that

• $x_i, x_j$ are the same person -> $d(x_i, x_j)$ is small
• $x_i, x_j$ are the different people -> $d(x_i, x_j)$ is large

How to train? Triplet loss

• Anchor $A$
• Positive $P$
• Negative $N$
• Want:
  • $d(A, P) \leq d(A, N)$
  • $d(A, P) - d(A, N) \leq 0$
  • This can be satisfied trivially with $d(*) = 0$.
  • To prevent trivial solution, require the difference larger than a margin. $d(A, P) - d(A, N) + \alpha \leq 0$.

End up with Triplet loss $\mathcal L(A, P, N) = max(d(A, P) - d(A, N) + \alpha, 0)$.