Fig. 1: Illustration of the key characteristics of the neural oscillation mechanism and the design of Rhythm-SNN.

a Neural oscillations spanning a wide range of frequencies have been observed across various brain regions, which play crucial roles in neural computation. Top Right: Neural oscillations operating at various frequencies enable the brain to synchronize and integrate information across diverse timescales. Bottom Left: Neural oscillations enhance energy efficiency by selectively activating distinct neural populations at specific phases of the oscillatory signal. Bottom Right: Neural oscillations promote pattern separation, allowing for robust decoding of the target signal from noisy inputs. b In the proposed Rhythm-SNN, neurons are modulated by oscillatory signals of different frequencies, which are represented by different colors. c Neuronal dynamics of rhythmic spiking neurons depicted in (b). The charging and firing behaviors of these neurons are influenced by the square wave modulation signals. Note that a constant input current is applied to these neurons in this illustration. d The unfolded computational graph of rhythmic spiking neurons is shown in (c). These neurons alternate periodically between `ON' and `OFF' states following neural modulation. During the `OFF' state, membrane potentials remain unchanged during forward propagation, thereby conserving energy. In backward propagation, gradients effectively propagate by skipping the `OFF' states, thus establishing a highway for gradient backpropagation through time.