Unveiling the Disparity in Perception: Neural Networks vs. Human Sensory Recognition

A study conducted by MIT neuroscientists has found that deep neural networks, while capable of identifying objects similar to human sensory systems, often produce unrecognizable or distorted images and sounds when prompted to generate stimuli similar to a given input. This suggests that neural networks develop their own unique invariances, diverging from human perceptual patterns. The researchers propose using adversarial training to make the models' generated stimuli more recognizable to humans, providing insights into evaluating models that mimic human sensory perceptions.