Categories vs Semantic Features: What shape the similarities people discern in photographs of objects?

In visual cognitive neuroscience, there are two main theories about the function ofthe ventral visual. One suggests that it serves to classify objects (classificationhypothesis); the other suggests that it generates intermediate representationsfrom which people can generate verbal descriptions, actions, and other kinds ofinformation (distributed semantic hypothesis). To adjudicate these, we trainedtwo deep convolutional AlexNet models on 330,000 images belonging to 86categories, representing the intersection of Ecoset images and the semanticnorms collected by the Leuven group. One model was trained to producecategory labels (classification hypothesis), the other to generate all of an item’ssemantic features (distributed semantic hypothesis). The two models learned verydifferent representational geometries throughout the network. We also estimatedthe human semantic structure of the 86 classes by using a triadic comparisontask. The representations acquired by the feature-generating model aligned betterwith human-perceived similarities amongst images, and better predicted humanjudgments in a triadic comparison task. The results thus support (distributedsemantic hypothesis).