Analog optical neural networks (ONNs) can reduce the energy of matrix-vector multiplication in neural network
inference below that of
digital electronics. However, realizing this promise remains challenging due to digital-to-analog (DAC) conversion—even at low bit precisions b, encoding 2b levels of digital weights and inputs into the analog domain involves power-hungry
electronics. Faced with similar challenges,
telecommunications uses complex-valued Quadrature-
Amplitude Modulation (
QAM).
QAM maximally exploits the
complex amplitude to provide a quadratic 0(N2)→0(N) energy saving over intensity-only modulation. QAMNet, an ONN with lower
energy consumption than existing ONNs, uses the complex nature of the amplitude of light with
QAM. QAMNet accelerates complex-valued
deep neural networks with accuracies indistinguishable from digital hardware. Compared to standard ONNs, QAMNet ONNs are (1) more accurate above moderate levels of total bit precision, (2) more accurate above low energy budgets, and (3) an optimal choice when hardware bit precision is limited.