A Reweighted Scheme to Improve the Representation of the Neural Autoregressive Distribution Estimator

Algorithm 1

A reweighted scheme for new feature.

	Given: An observation , weight matrix , and bias from the trained NADE
	Output: New feature corresponds to the observation
(1)	Sum the activations of all hidden layers and normalize it to get a vector using equation (13)
(2)	For each dimension j in vector , compare with the thresholds and set the corresponding weight to using equation (14)
(3)	Normalize the weight scores using equations (15) and (16)
(4)	Modify using equations (17) and (18)
(5)	Compute the new feature using equation (19)
(6)	Choose the best feature by validation.