MIT Researchers Discover a Training Method to Make Computer Vision Models more Like Humans

MIT researchers have found that a specific training technique called adversarial training can help computer vision models learn more stable and predictable visual representations, similar to those that humans learn using a biological property called perceptual straightening. The researchers hope that this will lead to the development of models that make more accurate predictions, which could improve the safety of autonomous vehicles. The study will be presented at the International Conference on Learning Representations. Researchers hope to create new training schemes that would explicitly give a model this property and dig deeper into adversarial training to understand why this process helps a model straighten.