標籤：

[圖集] mnist

05-12

Input: 28*28=784

Output: 10*1=10

Average of different classes $E(x_i),iin[0,9]$

Eigenvectors of expected crossproduct matrix $E(xx^T)$ (784*784).

Eigenvectors of expected crossproduct matrix $E(x_3x_3^T)$ (784*784).

Fitted filters

Degenerate distribution for label (only "3" visible in training)

("3" and "6" visible in training)

Averaged "6" - Averaged "3"

Fully randomised labels

推薦閱讀：

※ML領域的生物進化論，進化策略圖文詳解
※實對稱矩陣的對角化
※Inception-v2/v3結構解析（原創）
※RNN Part 3-Back Propagation Through Time and Vanishing Gradients(BPTT演算法和梯度消失)

TAG:神經網路 |