Classifying images utilizing generative-discriminative feature representations

A system comprising:

• one or more memory devices comprising:
• a generative neural network;
• and
• a classification neural network;
• and
• one or more computing devices configured to cause the system to:
• generate a reconstructed image from an input image by utilizing the generative neural network to:
• encode the input image into a latent space to generate a first encoding;
• and
• decode the first encoding to regenerate the input image;
• generate an input combination by concatenating the reconstructed image and the input image;
• and
• determine a classification for the input image that indicates whether the input image depicts content corresponding to one of a set of known classifications or unknown content corresponding to an open-set category by utilizing the classification neural network to:
• encode the input combination into a feature space to generate a second encoding;
• and
• determine a plurality of classification scores based on the second encoding.

A computer-implemented method comprising:

• encoding features from an input image in a vector embedding;
• generating, utilizing a generative neural network, a reconstructed image from an input image by decoding the vector embedding to pictorially represent the features in the reconstructed image;
• generating an input combination by combining the reconstructed image and the input image;
• determining a plurality of classification scores for the input image by analyzing encoded features of the input combination utilizing a classification neural network;
• and
• determining a classification for the input image based on the plurality of classification scores that indicates whether the input image depicts content corresponding to a known classification or unknown content corresponding to an open-set category.