Figure 1.

Representational spaces for object semantics. (a) The final layer of a visual DNN¹⁰ represents object in terms of 1,000 labels, and activations for different objects on the final layer can be visualized as points in the 1000-dimensional space spanned by the labels (for illustration, only 3 of the 1,000 dimensions are shown). All accurately-labeled objects will have orthogonal vectors in this space (grey lines) and so will be maximally dissimilar in this space, irrespective of the objects’ semantics. (b) A representation of objects in terms of their semantic feature vectors (e.g. using concept property norms). Each object is represented as a pattern of activation over semantic feature units (for illustration, only 3 dimensions are shown). Semantically similar objects, such as tangerine and banana will have high activations for shared semantic features and thus will be closer together in this space than to other semantically unrelated concepts (e.g. motorcycle). Object images reprinted with permission from Hemera Photo Objects.