Domain adaptions

Problem: Model we train on one training set does not perform well on another data with other intensity character

CNNs are not robust against domain shift

Domain: sinset of samples showing similar variations

Source domain: subset which the models are trained

Target domain: subset wgich models are applied

Domain shift is target domain is not subset of source domain —> algorithms are not generalizing

Seperate training for each domain with separate training set and final model

Inference at each domain with specific model

+ If engough labels present —> best performing model

- requires a lot of data in every domain

Inital training on the source domain with many samples

FIne-tune model at the target domain

Moel training starts with optimal source domain parameters

+ Only few labeled samples at the target domain

- Target domain still requires labels for fine-tuning

Reduce the number of needed samples by freezing layers

Change only the batch norm layers in the network

Training with source domain an unlabeled target domain

Supervise loss: task specific, use labels of source domain

Adversarial loss: measures distance between feature distribution of samples from source and target domains

+ Does not require any labeled images at target domain

- Need of transporting eacg target domain

Transformations of the data sample = new sample

Geometric transformations: affine/elastic transformations

Intensity transformations: blur, noise, gamma, etc.

Tackle lack of data: mimic new domains

Stacked transformations:

Imag quality transformations: sharpness, blurry, noise

Image appereance transformations: brightness, contrast

Through the augmentation the source domain is expanded hoping that the expanded set overlaps with the target domain

Design flexible models generalizing well on new domains

Idea: Split dataset to many small ones and learn to lern

Introduce meta train step

Parameter update step:

Entropy: Uncertainty of a prediction

Assumption: pushing confidence of predictions higher —> leads to goo domain generalizations

+ Easy to implement, no prior information, applicable to different tasks

- Strong assumption, many test samples needed

Autoencode loss of input, output and intermediate features

AE: map from image to image itself x = D(E(x)) where D decodes and e encodes

Idea: Reconstruct input, intermediate and output features with AE on source domain —> new images are converted to source domaine-like images —> right predictions

+ Applicable to different tasks, strong prior model

- Input, intermediate features can be prone to dom shifts

Denoising auto encoder: map from noisy version to itself x=D(E(x_noise))

Same idea as with AE: getting source like images

+ Accurate segmentation results, model not affected by domain shift

- Specific to segmentation task