FIGURE 3.

Examples of reconstructions obtained using causes inferred by the trained model without receptive fields. Each column represents an example of a sensory input. The three leftmost images represent novel stimuli from object classes used in training whereas other images are from object classes not used in training. The top row shows the novel sensory input that was presented to the network to allow it to construct latent representations across the areas. The second to fifth rows show the reconstructions of the sensory input obtained using the latent representations in the corresponding areas of the model. It can be observed that the reconstructed sensory input faithfully reproduces the novel originals, although the lower areas regenerate the inputs more sharply.