Training Machines To Learn Like Humans

Researchers at MIT have developed a technique to help computer vision models learn more perceptually straight representations, similar to how humans process visual information. This technique, called adversarial training, makes the model less sensitive to small errors added to the image and improves the perceptual linearity of the model. The team also found that the task the model was trained and run on affects its perceptual accuracy. By better understanding perceptual accuracy in computer vision, researchers hope to develop models that make more accurate predictions.