Eigenvalue communication is a promising technology that can overcome capacity limitations caused by the Kerr nonlinearity in optical fibers. In this letter, we propose an artificial neural network-based demodulation method for optical eigenvalue-modulated signals using on-off encoding. The proposed method has a power margin improvement of 9.2 dB at the bit error rate of 3.8 × 10−3 compared with the conventional inverse scattering transform-based demodulation method in numerical simulations.