Siamese network vs fully-connected network

[karmic olive]

What is the difference between concatenating 2 vectors and passing them through a fully-connected network, then a sigmoid layer; and passing each vector through a Siamese network that maps them into a latent space and then for the distance between them to be calculated?

[pure linden]

basically, if you just concatenate input through one model together you can't compare the output between the different inputs anymore. when you run each through the model separately the contrastive loss function optimizes the model to find the features that are different between classes

[karmic olive]

Is there a class of classification problem where Siamese networks are worse than fully connected?

[pure linden]

well, siamese networks aren't technically doing a classification task. you are more creating an embedding and fully connected network to interpret the differences. a fully connected classifier is just a very different method

my experience with them is normally in computer vision and you would normally use the same type of pretrained cnn for both tasks just setup differently

[karmic olive]

Is it possible to have data that can't be embedded well and would be better to be passed through a fc?

What about multi-modal data?

[pure linden]

if you have a finite number of classes it would probably be more efficient to just directly train a classifier. using contrastive loss to create an embedding is like more complicated

[hushed mesa]

It is not really that you can't compare the output, if you concat the input and them pass through a FC model and train like this, you can't use the model to project a single input anymore.
But you can do that if you use a model that takes a single input at a time and use this one two times (normal siamese networks).