Adversarial_Discriminative_Domain_Adaptation.md

Adversarial Discriminative Domain Adaptation

What's the main idea and/or problem to solve?

• Domain shift/adaptation. If we train a network to classify digits of MNIST, for example, how do we also get it to work on USPS digits? That's the domain shift problem, and it's intuitively very hard but also quite important to know how to train networks to generalize. Or how we make the "latent feature space" of the network to be domain invariant.

• More formal problem statement: we are given source images and labels (X_s,Y_s) drawn from the source domain distribution p_s(x,y), as well as target images X_t from the target distribution p_t(x,y) but where we have NO LABELS. The goal? Learn a target representation M_t and classifier C_t that can correctly classify the target images into one of K categories at test time.

Why does their method work?

• It helps to think of a neural network that does classification as two parts: a feature embedding followed by the actual classifier (see my comments & bullet points above). So perhaps the "feature embedding" consists of the convolutional layers of the net, and then the "classifier" is the fully connected part.

• So, what to do? Learn a feature embedding for both the source and target images that result in similar embeddings. Then the trained classifier portion for the source can simply be applied on the target. That makes sense.

• Minimize source and target representation distances between alternating minimizations of the two steps:

  • A domain discriminator classifies whether a data point is drawn from the source or target domain. Not the labels --- the data points. And BTW, these data points are fed through the feature embedding and that's ultimately what we use for the cross entropy loss. Yes, that makes sense.

  • Source and target mappings (i.e. the M_s and M_t terms) are optimized according to a constrained adversarial objective, which varies among different instantiations of their framework. In general, M_s and M_t have a similar neural network architecture. They may or may not share weights, which is one way to enforce equality (but then it might not make M_t "discriminative enough" among the target distribution's categories).

• I think I understand why this works. The adversarial objective is trying to "overpower" the discriminator so that it creates similar feature embeddings for both the source and target. The discriminator gives a good gradient.

• They cast their algorithm definition as picking among three categories: generative vs discriminative base model, tie/untie weights, and which adversarial objective. The utility of this is subject to some questioning, though.

Could I implement this?

• I think I can implement the discriminator objective. That's typically the easier part with GAN-like formulations anyway.

• The source M_s is learned the normal way with cross entropy loss. And during adversarial training, it is fixed. I think this makes sense, since in normal GANs we have a fixed true distribution.

• Now what about M_t?? It is pre-trained with weights initialized from M_s. This makes sense; it's like how in GAIL we should initialize the generator from behavior cloning. It's trained with the adversarial loss, and both M_t and the discriminator are trained in this stage.

Yes, I think I could implement this given enough time. Figure 3 is very helpful.