Fig. 3

The ML component schematic. Conv1: A 14 channel, \(5 \times 5\) kernel, 1-stride, 1-padding convolution is applied to the single channel input. This is followed by an ELU activation, and a \(2 \times 2\), stride 2 max pooling step. Conv2: A 15 channel, \(3 \times 3\) kernel, 1-stride, 1-padding convolutional layer. This is followed by an ELU activation, and another max pooling step. An average pooling is performed over each channel, and the RSVD-determined eigenvalue is appended to this vector of channel averages. This passes through width-16 ELU-activation dense layers, FC3 and FC4, before passing through a Sigmoid activation for binary classification of eigenpatch utility.