Image generation model based on log-likelihood

A system for training an image generation model, the image generation model being configured to generate an image from a latent feature representation, the system comprising:

• a data interface configured to access model data representing parameters of the image generation model, and training data representing a training dataset of multiple training images;
• a processor subsystem configured to train the image generation model using a log-likelihood optimization and output the trained image generation model, wherein:
• the training includes selecting a training image of the training images, and determining a log-likelihood of the training image being generated according to the image generation model,
• the image generation model includes a transformation configured to determine a discrete feature from a continuous feature vector, the transformation being an argmax transformation configured to compute a discrete index feature indicating an index of a feature of the continuous feature vector with an extreme value,
• determining the log-likelihood includes obtaining a value of the index feature for the training image, sampling values of the continuous feature vector given the value of the index feature according to a stochastic inverse transformation of the argmax transformation, and determining a likelihood contribution for the argmax transformation based on a probability that the stochastic inverse transformation generates values of the continuous feature vector given the value of the index feature.

A non-transitory computer-readable medium on which is stored instructions for training an image generation model, the image generation model being configured to generate an image from a latent feature representation, the instructions, when executed by a computer, causing the computer to perform the following steps:

• accessing model data representing parameters of the image generation model, and training data representing a training dataset of multiple training images;
• training the image generation model using a log-likelihood optimization, wherein:
• the training includes selecting a training image of the training images, and determining a log-likelihood of the training image being generated according to the image generation model,
• the image generation model includes a transformation configured to determine a discrete feature from a continuous feature vector, the transformation being an argmax transformation configured to compute a discrete index feature indicating an index of a feature of the continuous feature vector with an extreme value,
• determining the log-likelihood includes obtaining a value of the index feature for the training image, sampling values of the continuous feature vector given the value of the index feature according to a stochastic inverse transformation of the argmax transformation, and determining a likelihood contribution of the argmax transformation for the log-likelihood based on a probability that the stochastic inverse transformation generates values of the continuous feature vector given the value of the index feature;
• outputting the trained image generation model.