Add to Favourites
To login click here
MIT researchers have discovered that a specific training technique, adversarial training, can enable certain types of computer vision models to learn more stable, predictable visual representations, which are more similar to those humans learn using a biological property known as perceptual straightening. The team also found that the task one trains a model to perform affects the perceptual straightness of the model, with models trained to perform abstract tasks learning more perceptually straight representations than those trained to perform more fine-grained tasks.
