Skip to main content

Table 3 The distribution of confidences

From: Understanding deep learning via backtracking and deconvolution

  Probability
Angry Contemptuous Disgusted Fearful Happy Neutral Sad Surprised
Case 1
 Before 6.36e−10 4.5e−7 6.71e−7 9.89e−1 2.59e−12 3.34e−3 7.27e−3 5.10e−4
 After 5.57e−5 2.71e−4 5.57e−5 1.22e−1 5.57e−5 5.79e−3 8.72e−1 5.57e−5
Case 2
 Before 1.92e−1 6.95e−1 2.67e−6 3.66e−10 6.52e−10 3.17e−5 1.13e−1 1.39e−12
 After 9.12e−1 4.56e−3 7.68e−7 7.45e−7 7.45e−7 1.72e−6 8.32e−2 7.45e−7
Case 3
 Before 2.70e−9 1.86e−7 7.16e−11 4.13e−1 1.26e−11 5.87e−1 2.41e−6 2.90e−9
 After 1.32e−9 3.01e−9 1.32e−9 9.99e−1 1.32e−9 5.53e−5 1.84e−9 1.32e−9
Case 4
 Before 9.75e−1 5.47e−6 4.95e−10 7.94e−12 2.09e−17 6.16e−9 2.46e−2 1.43e−13
 After 1.54e−2 6.97e−5 1.04e−7 1.04e−7 1.04e−7 2.99e−6 9.85e−1 1.04e−7
Case 5
 Before 6.89e−9 2.55e−6 9.54e−1 4.30e−2 8.70e−10 2.52e−3 4.87e−5 3.81e−7
 After 9.58e−5 9.74e−5 1.14e−2 8.08e−1 9.58e−5 1.66e−1 1.35e−2 9.58e−5
Case 6
 Before 9.64e−2 7.11e−4 1.92e−10 2.00e−5 1.77e−14 2.17e−4 9.03e−1 4.44e−15
 After 9.87e−1 7.78e−4 1.21e−8 1.26e−8 1.21e−8 4.12e−7 1.18e−2 1.21e−8
Case 7
 Before 2.08e−13 9.97e−9 5.66e−16 3.36e−1 1.29e−14 1.53e−1 2.37e−9 5.11e−1
 After 5.13e−7 5.13e−7 5.13e−7 8.98e−1 5.13e−7 9.85e−2 5.13e−7 3.25e−3
Case 8
 Before 2.81e−12 1.02e−10 1.74e−1 1.65e−17 8.26e−1 8.52e−15 4.41e−14 1.39e−18
 After 6.55e−10 6.55e−10 9.79e−1 6.55e−10 2.13e−2 6.55e−10 6.55e−10 6.55e−10
  1. Italic values indicate probabilities of the miss-classified classes (before the suppression) and the probabilities of the correct classes (after the suppression)