Domain Adaptation

Domain Adaptation

Domain adaptation is a field of computer vision, where the goal is to train a neural network on a source dataset and secure/achieve a good accuracy on the target dataset which is significantly different from the source dataset.

e.g. From CAD models to real images:

Maximum Mean Discrepancy (MMD)

Maximum Mean Discrepancy Loss

Suppose the mean of source and target features is given by where is the feature of the th source image, is the number of source features, is the feature of the th target image, is the number of target features. Then the maximum mean discrepancy is defined as

Maximum Mean Discrepancy Method

Suppose we have a pre-trained feature extraction network with be the normal classification loss that extracts features from both source and target dataset. In each training mini-batch, we minimize the following overall loss: where controls the balance between the two losses.

Correlation Alignment

CORAL Loss

The CORAL loss is defined as the distance between the second-order statistics (covariances) of the source and target features: where is the feature dimension, and and represent covariance matrix of the source and target domains respectively.

Correlation Alignment

Suppose we have a pre-trained feature extraction network with be the normal classification loss that extracts features from both source and target dataset. In each training mini-batch, we minimize the following overall loss: where where is the number of layers that apply this loss.

Adversarial Alignment

Domain Classification Loss and Domain Confusion Loss

Domain classification loss is defined as Domain confusion loss is defined aswhere , indicate source, target domain samples. is the parameters of feature extraction network, is the parameters of domain classifier. means the predicted domain of an input sample, and is the ground truth domain of this sample. 𝑝 is the softmax of the domain classifier activation:

Adversarial Alignment

Perform iterative updates for the following two objectives given the fixed parameters from the previous iteration: