Fig 1.
Schematic description of the computing network.
The N inertial masses (circles) arranged in a chain are coupled to neighbors by linear springs, and to a substrate by linear or non-linear springs, with damping. A harmonic forcing, with amplitude possibly modulated by couplings to the input signal u(t), is imposed on the masses.
Fig 2.
Computing parity functions with a network of coupled anharmonic oscillators.
Driving term u(t), P3(t), P4(t), P5(t) (top to bottom, shifted vertically for clarity). Green curves correspond to the training phase (t < 0). For t > 0, red curves correspond to the target functions, while blue curves correspond to the network outputs.
Fig 3.
Words classification benchmark results.
The color-scale indicates the probability that a digit presented to the device (columns) is classified to a certain value (lines) by the oscillators network. The numbers at the top of each column indicate the success probability (prediction matches the actual digit), estimated with an uncertainty of ±1% (95% confidence level).
Fig 4.
Global tuning of network parameters.
Success probability (P) for the three parity functions (P3, P4, P5, left to right) as the period T of the input binary signal and the amplitude A of the oscillator drive are varied globally for the whole network.
Fig 5.
Robustness of the parity benchmark for pre-training variations.
Variations in the success probability (P) for the three parity functions (blue: P3, green: P4 and red: P5) as the relative variation σ is increased, for perturbations introduced before the training of the network is performed. Error bars are computed at the 95% confidence level.
Fig 6.
Robustness of the parity benchmark for post-training variations.
Variations in the success probability (P) for the three parity functions (blue: P3, green: P4 and red: P5) as the relative variation σ is increased, for perturbations introduced after the training of the network is performed. Error bars are computed at the 95% confidence level.