Abstract
We demonstrate that neural networks can be FLOP-efficient integrators ofone-dimensional oscillatory integrands. We train a feed-forward neural networkto compute integrals of highly oscillatory 1D functions. The training set is aparametric combination of functions with varying characters and oscillatorybehavior degrees. Numerical examples show that these networks areFLOP-efficient for sufficiently oscillatory integrands with an average FLOPgain of 1000 FLOPs. The network calculates oscillatory integrals better thantraditional quadrature methods under the same computational budget or number offloating point operations. We find that feed-forward networks of 5 hiddenlayers are satisfactory for a relative accuracy of 0.001. The computationalburden of inference of the neural network is relatively small, even compared toinner-product pattern quadrature rules. We postulate that our result followsfrom learning latent patterns in the oscillatory integrands that are otherwiseopaque to traditional numerical integrators.