Researchers at Tel-Aviv University have created a method to solve regression tasks with little supervision effectively. The researchers noted that GANs are great for encoding semantic information into the latent space, and they see this as important in modern generative frameworks. This means smooth linear directions, which affect image attributes disentangled. These directions have been used in GAN-based image editing for years. The research team found that these directions are not only linear, but the magnitude of change induced on respective attributes also follows a relatively flat pattern no matter how far they go along them. This new method uses this observation to turn any pre-trained GAN into a regression model by leveraging as few as two labelled samples instead of many more previously required.
In this new approach to GAN usage, the research team focused on extracting information from latent spaces in modern GANs. They took their cues from StyleGAN’s semantic space and applied it across various generative tasks for downstream processing.
Researchers have found that semantic properties can be used to train regression models. They accurately predicted the magnitude of a particular attribute in an image by measuring its distance from a separating hyperplane induced by matching linear latent directions. Researchers have found that latent-space distances can serve as regression scores for applications where no conventional units are required or exist. However, they noticed an interesting phenomenon – typically, these distances were also approximately linear with respect to the magnitude of the semantic attribute. Therefore, they calibrated them so that real world predictions could be made by using linear regression and only needing two labelled samples which would help in data domains where quality supervision is prohibitively difficult to acquire.
Outline of the summary:
• ‘The observation that commonly, latent-space distances are approximately linearly correlated with the magnitude of semantic properties in an image’.
• ‘A scheme for converting a pretrained generator and a semantic latent-direction into a state-of-the-art few-shot regressive model’.
• ‘A new approach to analysing layer-importance and mapping semantic distances between the latent spaces of a GAN’.