Overview

• MIT researchers demonstrated that the ventral visual stream is not solely optimized for object recognition but also processes spatial features like location, orientation, and distance.
• Convolutional neural networks (CNNs) trained on spatial tasks showed similar levels of neuro-alignment with the ventral stream as models trained on object recognition tasks.
• A synthetic dataset featuring objects in varied orientations and backgrounds was used to train CNN models to extract spatial features.
• Analysis revealed that early to middle layers of CNNs trained on different tasks develop nearly indistinguishable representations, suggesting shared foundational processing stages.
• The findings challenge traditional views of the ventral stream's function, prompting a reassessment of its role in visual processing and its implications for neuroscience and AI research.