Example of our CNN architecture for the frequency–time array of a dedispersed FRB. The input image is a simulated strong, scattered burst, artificially bright for the sake of this example. The input data are convolved with 32 different convolutional kernels, the results of which are fed through a nonlinear ReLu function, then reduced in a max-pooling layer. Sixty-four more kernels are then applied to the binned arrays, and pooling is done again by taking the maximum value in each 2 × 2 bin. Each successive layer in the network’s feature extraction component is meant to discover features at higher levels of abstraction. We have included in this figure the real activations of a trained model for this input image. These give an idea of what the neural network’s classification is based on and can help with the “black-box” problem. After the fully connected layers, a probability is assigned to the trigger’s likelihood of being an FRB.