In recent years, there have been enormous improvements in event reconstruction and signal identification for neutrino experiments via deep learning methods. Despite the success in accuracy of current ML algorithms, the power-efficiency of such methods calls out for improvement for any hope at realizing real time machine learning data processing. Solutions are two-fold: either resorting to software, using a faster algorithm and gaining absolute speed under the same power consumption, or bringing into play a hardware alternative that consumes less energy in performing the same operations.
In this work, we turn to the hardware alternative, presenting the first attempt at accelerating deep learning methods for muon event reconstruction on Tensor Processing Units (TPU’s). We use an LSTM-based recursive neutral network with CNN-based data encoding that is compatible with Google TPU hardware requirements. We show that this algorithm is capable of achieving similar angular resolution in reconstruction compared to State of the Art ML peers, and reaches a performance per watt of 100 Hz/Watt on a TPU accelerator. This opens up an entire world of chances at integrating machine learning capacity into detectors and electronics that go deep into even the most power-restricted environments.