Regularizers

Various regularization techniques are implemented regarding the design of optical coding elements. These regularizations force certain physical constraints on the learning parameters. Also, there are regularizers that promotes certain behaviour on the coded measurements to imporve the task perfomance.

The regularizers are used in the optimization problem of the form:

\[\{\learnedOptics^*,\theta^*\} = \arg \min_{\learnedOptics,\theta} \sum_{p=1}^{P}\mathcal{L}(\mathbf{x}_p,\reconnet( \forwardLinear_{\learnedOptics}(\mathbf{x}_p))) + \lambda \mathcal{R}_1(\phi) + \mu \mathcal{R}_2(\forwardLinear_\learnedOptics (\mathbf{x}))\]

where \(\mathcal{L}\) is the task loss function, \(\mathcal{R}_1\) is a regularizer over the coding elements, and \(\mathcal{R}_2\) is a regularizer over the coded measurements \(\lambda\) and \(\mu\) are the regularization weights, and \(P\) is the number of samples in the training dataset.

List of Regularizers

colibri.regularizers.Binary	Binary Regularization for Coded Aperture Design.
colibri.regularizers.Correlation	Correlation Regularization for the outputs of optical layers.
colibri.regularizers.Transmittance	Transmittance Regularization for Coded Apeuture Design.
colibri.regularizers.KLGaussian	KL Divergence Regularization for Gaussian Distributions.
colibri.regularizers.MinVariance	Minimum Variance Regularization.