O'Reilly logo
  • NewRelic SyntheticsTest thinks this is interesting:

We want to be judicious with choosing a filter count for a convolutional layer because computing the activations of a single convolutional filter is more expensive than traditional multilayer neural network layers. In DL4J, we configure the filter count with the nOut(int) option in the configuration.

As we go further along in the progression of layers in a CNN, the activation map size decreases. The layers closer to the input layer will tend to have fewer filters. As we move toward the output layer in a CNN, we tend to see more filters.

From

Cover of Deep Learning

Note

TEST TEST PUBLIC NOTE -- We want to be judicious with choosing a filter count for a convolutional layer because computing the activations of a single convolutional filter is more expensive than traditional multilayer neural network layers. In DL4J, we configure the filter count with the nOut(int) option in the configuration.

As we go further along in the progression of layers in a CNN, the activation map size decreases. The layers closer to the input layer will tend to have fewer filters. As we move toward the output layer in a CNN, we tend to see more filters.