Similarity Learning and Generalization with Limited Data: A Reservoir Computing Approach

Figure 8

Graphing probabilities of the reservoir output () versus the image pair number , for image pairs that are rotated (a, d) 2 combination: rotated and blurred (b, e); 3 combination: noise, blurred, and zoomed (c, f), for single and dual reservoir, respectively. The fractions correct, where classification is considered to be correct if the predicted maximum probability labels are the transformations applied to the test image pair (shown on top left of each panel), are , , , , , and for (a, b, c, d, e, f), respectively. =0.5; reservoir size=1000. Training digits: 0-5; testing digits: 6-9. Training size: 250 pairs.

(a)

(b)

(c)

(d)

(e)

(f)