Machine learning model development

A method of machine learning model development, the method comprising:

• building an autoencoder including an encoder trained to map an input into a latent representation, and a decoder trained to map the latent representation to a reconstruction of the input;
• building an artificial neural network classifier including the encoder of the autoencoder, and a classification layer partially trained to perform a classification in which one of a number of classes to which the input belongs is predicted based on the latent representation;
• applying neural network inversion to the classification layer to find inverted latent representations within a decision boundary between classes of the number of classes in which a result of the classification is ambiguous;
• obtaining inverted inputs from the inverted latent representations;
• labeling each inverted input of the inverted inputs with one of the number of classes that is its ground truth, and thereby producing added training data for the classification;
• and
• further training the classification layer using the added training data.

A computer-readable storage medium for machine learning model development, the computer-readable storage medium being non-transitory and having computer-readable program code stored therein that, in response to execution by processing circuitry, causes an apparatus to at least:

• build an autoencoder including an encoder trained to map an input into a latent representation, and a decoder trained to map the latent representation to a reconstruction of the input;
• build an artificial neural network classifier including the encoder of the autoencoder, and a classification layer partially trained to perform a classification in which one of a number of classes to which the input belongs is predicted based on the latent representation;
• apply neural network inversion to the classification layer to find inverted latent representations within a decision boundary between classes of the number of classes in which a result of the classification is ambiguous;
• obtain inverted inputs from the inverted latent representations;
• label each inverted input of the inverted inputs with one of the number of classes that is its ground truth, and thereby produce added training data for the classification;
• and
• further train the classification layer using the added training data.