Straightening Out Ai: How Mit Researchers Bridge The Gap Between Human And Machine Vision

MIT researchers have discovered that adversarial training can improve perceptual straightness in computer vision models, making them more similar to human visual processing and enabling better prediction of object movements. Adversarially trained models are more robust, retaining a stable representation of objects despite slight changes in images. The researchers aim to use their findings to create new training schemes and further investigate why adversarial training helps models mimic human perception.