| CPC G06N 3/045 (2023.01) [G06N 3/088 (2013.01)] | 20 Claims |
|
1. A method, comprising:
receiving, at a deep machine learning model, a training dataset having a pair of samples, including a first pair and a second pair, the first pair including a first training sample and a first attribute and the second pair including a second training sample and a second attribute;
training an encoder of the deep machine learning model, using a contrastive objective, to generate a first latent vector in a latent space for the first pair and a second latent vector in the latent space for the second pair;
training the latent space using a smoothing objective;
reconstructing, using a decoder of the deep machine learning model, a reconstructed pair of samples including a reconstructed first pair from the first latent vector and a reconstructed second pair from the second latent vector;
determining a reconstruction loss using the pair of samples and the reconstructed pair of samples;
training the encoder, using a consistency learning objective, to rank the reconstructed first pair and the reconstructed second pair;
updating, at least one parameter in the deep machine learning model, based on at least one of the contrastive objective, smoothing objective, reconstruction loss, or consistency learning objective;
determining that a reconstructed first sample in the reconstructed first pair has a higher rank than a second reconstructed training sample in the reconstructed second pair; and
discarding the second pair based on the determining.
|