Skip to main content
. 2019 Mar 18;29(6):979–990.e4. doi: 10.1016/j.cub.2019.01.077

Figure 5.

Figure 5

Size Invariance and Grid Cell Lesions

(A) Top: Illustration of the relationship between distances in “neural space” (on grid patterns) and visual space. Panels 2–4: Scaled down (50%) stimuli within the presentation frame of default size. Scaling the gain between the displacement vector on the grid pattern and the oculo-motor output uniformly for all saccades allows the model to sample all features of downscaled stimuli, leading to successful recognition irrespective of size. Saccade sequences (red arrows) are superimposed on stimuli (left). Cyan circles indicate the centers of all encoded local features. With each sampled feature, a firing rate of the corresponding identity cell is incremented (right).

(B) Histogram of the number of saccades (from last reset) necessary for recognition across all stimuli.

(C) Upon disconnecting grid cells from the model, it is assumed bottom-up attention will select among possible targets randomly. Panel 1: The next predicted feature (red circle) and the foveated feature rarely match, leading to poor evidence accumulation. Panel 2: By chance the next predicted feature (red circle) can be selected for foveation. Panel 3: The absence of confirmatory predictions (and thus facilitation among feature label cells) increases the number of false hypotheses of stimulus identity, leading to predicted features that are not present in the stimulus being viewed (purple circle). The behavior in panels 1 and 3 reduces recognition performance, while that in panel 2 contributes to residual recognition ability.

(D) Cyan circles indicate memorized features. Adding additional potential targets for bottom-up attention (i.e., distractors, filled green circles) exacerbates the performance drop.

(E–G) The average number of fixations from the last reset across conditions (E). The average number of fixations including resets (F). Recognition and failure rates (no recognition within 10 resets, “not rec.”) for all tested conditions (G). Default condition, blue; white noise and real-world occlusions, red and yellow, respectively; real-world occlussions limited to one consecutive saccade towards the occlusion, purple; 50% shrunken stimuli, green; grid cell lesion with and without distractors, light blue and dark red, respectively.

Image credit: B.B., P.A.: used with permission; Mr. Spock, Ai Wei Wei, Kiwi, bathroom: public domain images.