Figure 9. Hallucinations with visual impairment restricted to the top half of the input.

(A) Example images. (B) Hallucination qualities during adaptation. Note that many of the corresponding decoded internal representations were not actually hallucinations, but rather matched shapes that were in the unimpaired half of the visual input. In particular, in the early phase of adaptation there are two clusters at low and high quality values. These correspond to void internal representations or veridical ones when shapes happened to lie completely in the impaired or healthy halves, respectively. The former then were gradually replaced with emerging hallucinations. (C) Distribution of hallucinated small and large shape categories across the image in the model with fully impaired input. Only hallucinations with quality greater than 0.85 were counted here. (D) As C, but for the model that underwent adaptation with only the top half damaged (displayed data then taken with fully blank images as input as to not be influenced by actual objects in the healthy region). Now, hallucinations were localised to the impaired region and favoured smaller shapes, which would ‘fit’ within that region.