Architecture of the one-dimensional convolutional neural network used to recognize signals in LArTPC waveforms.
from: Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment Volume 1028, 1 April 2022, 166371, “Extracting low energy signals from raw LArTPC waveforms using deep learning techniques — A proof of concept – ScienceDirect “
The raw LArTPC waveform is presented to the input of the network (on the left in the diagram) which consists of three one-dimensional convolutional layers (Conv1Ds). To reduce the size of the input feature map, a pooling layer follows each Conv1D layer. The function’s output is bounded between 0 and 1, interpreted as the probability that a signal is within the waveform. There are around 20K trainable parameters in this approach, compared to the millions in typical designs.