Oscillatory Recurrent Gated Neural Integrator Circuits for Signal Processing

The patent changes the fundamental parameters of the neural network by introducing complex-valued weights and oscillatory basis functions instead of real-valued weights and piecewise constant activations. This allows the network to naturally represent and process oscillatory signals with time-varying amplitudes, frequencies, and phases while maintaining the ability to capture long-term dependencies.

The patent segments the complex LSTM functionality into simpler components by using a fixed oscillatory basis function set combined with learnable complex weights. This decomposition separates the temporal modeling (handled by the oscillatory basis functions) from the parameter learning (handled by the complex weights), making the system easier to design and more interpretable while maintaining long-term dependency representation.

The patent substitutes the digital computational mechanism with an analog electrical circuit implementation. The complex-valued computations are performed using physical electrical components (resistors, capacitors, operational amplifiers) that naturally handle complex arithmetic through their electrical properties, thereby achieving energy efficiency comparable to analog systems while maintaining the full computational capability of the recurrent neural network.