Reading this paper on one-shot learning "Siamese Neural Networks for One-shot Image Recognition" I was introduced to the idea of Siamese Neural Networks.

What I did not fully grasp was what they meant by this line:

This objective is combined with standard
backpropagation algorithm, where the gradient is additive
across the twin networks due to the tied weights.

Firstly, how exactly are they tied? Bellow, I believe I've provided the formula by which they compute the gradient. T is the epoch, $\mu_j$ is the momentum, $\lambda_j$ the regularization, $\eta_j$ the learning rate, $w_{kj}$ I believe to be the weight between neuron k and in one layer and j in another but correct me if I'm wrong.

\begin{equation}\begin{array}{c} \mathbf{w}_{k j}^{(T)}\left(x_{1}^{(i)}, x_{2}^{(i)}\right)=\mathbf{w}_{k j}^{(T)}+\Delta \mathbf{w}_{k j}^{(T)}\left(x_{1}^{(i)}, x_{2}^{(i)}\right)+2 \lambda_{j}\left|\mathbf{w}_{k j}\right| \\ \Delta \mathbf{w}_{k j}^{(T)}\left(x_{1}^{(i)}, x_{2}^{(i)}\right)=-\eta_{j} \nabla w_{k j}^{(T)}+\mu_{j} \Delta \mathbf{w}_{k j}^{(T-1)} \end{array}\end{equation}

My other question is why this is even desirable? Why not just reuse the same network twice? Or perhaps the two networks will be identical after training? If the networks are identical after training, why would you set it up like this? What benefits does it have?

  • $\begingroup$ Have a look at this post: datascience.stackexchange.com/a/43820/98414 which should provide the intuition behind it. $\endgroup$ Jun 6, 2020 at 13:21
  • $\begingroup$ Thanks for the link! In a Siamese network, the weights are identical on both branches. In the paper I was reading the weights are kept similar by computing the combined gradient for both sides and using that. Though I'm not sure how they do with initialization of the network. If both branches are initialized to different values the trained branches should differ (but probably not to much). $\endgroup$ Jun 7, 2020 at 11:32
  • $\begingroup$ I doubt they are initialized differently. They should share the same weights. $\endgroup$ Jun 7, 2020 at 12:04

1 Answer 1


You could have a single network, feed both inputs separately, compute the distance/loss, then perform backpropagation.


(as in the cited paper) you could initialize a network, and then create a parallel twin of that network. Because both networks see the same loss, they will remain identical after backpropagation.

The paper is explained in the form of the latter. In fact, I've always seen it explained in the form of the latter.

I can tell you in terms of implementation, its done using the former. As a matter of fact, this is directly from the GitHub for the cited paper.

enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.