Learning Domain-Independent Deep Representations by Mutual Information Minimization

Algorithm 1

Iterative learning algorithm of MMITR.

	Input: Training set of L classes and unlabeled data points ;
	Input: Domain indicators of training points ;
	Input: Tradeoff parameters and ;
	Input: Maximum number of iterations, η;
	Input: Objective value threshold, ε.
	Initialize iteration indicator .
	Initialize model parameters and objective value .
	while or objective value do
	E-step: Update the inter- and intraclass affinities for each data point, according to equations (5) and (6).
	M-step: Iterating the ADMM updating steps.
	for do
	Update the domain transfer representations by iterating the gradient descent steps of equation (27).
	Update the CNN model parameters by iterating the backprorogation steps of equation (30).
	Update the dual variables by iterating the gradient ascent steps of equation (31).
	endfor
	.
	endwhile
	Output: W and .