Skip to main content
. 2019 Apr 2;21(4):357. doi: 10.3390/e21040357

Figure 2.

Figure 2

On the MNIST dataset, Alexnet after 50 rounds of training outputs. (a) One sample from MNIST handwritten digit database, (b) standard probability (temperature = 1) and (c) soft-label probability (temperature = 3) for input. Soft labels can express the class-similarity relation more comprehensively.