Row 1.
In this experiment I started with the picture in 1(a). (Note that this picture does not  have to be an exact replica of a training image). Then I deleted some sections of  the image to get the image in 1(b). I then asked the system to recognize the image and generate a predicted image. The system correctly recognized the image and filled in the missing portions.

Row2.
I started with the same image as in 1(a) (replicated in 2(a)). I then deleted  out a few more sections of the image to obtain the one  in figure 2(b). The image in 2(b) was given to the system for recognition and filling in. The system now recognized the image as the one shown in figure 2(c) and did the filling-in/prediction according to that recognition.

This demonstrates that the system uses the global context to make predictions to fill in missing details.