Researchers have successfully demonstrated the behavior of simulated analog opto-electronic neuromorphic devices, utilizing tools previously developed for modeling these systems. The devices, modeled using VerilogA and compatible with SPICE models and Spectre simulators, are crucial for simultaneously simulating photonic and electronic effects, essential due to the fundamentally opto-electronic nonlinearity at play.
Simulations show that the timing and intensity of input pulses significantly alter the output response of the resonator neuron within the device. Specifically, the device reacts differently to single or double pulses depending on the spacing between the pulses in relation to the resonant period. Two pulses spaced half a resonant period apart can cancel each other out, preventing the system from reaching its threshold. Conversely, pulses spaced a full resonant period apart amplify the output signal, pushing the system past the threshold. This timing sensitivity allows the neurons to encode temporal information from inputs, potentially beneficial for tasks requiring coincidence detection and timing-based learning mechanisms.
The neuron’s behavior also varies with pump power. At lower pump power, the neuron is monostable, returning to a resting state after excitation. However, at higher pump power, the neuron exhibits bistability, maintaining a spiking state even without input. Returning to a resting state in this bistable regime necessitates a precisely timed pulse.
Further experiments revealed that the neuron can be excited by strong inhibitory signals due to the nature of its threshold as an unstable limit cycle. This is significant because the device outputs negative spikes relative to its DC operating point.
The research also explored the neuron’s response to constant excitation. Under sufficient constant input power, the neuron enters a self-pulsating state. Varying the input power demonstrated that the device operates as a Class 2 excitable system. This classification is characterized by an abrupt shift in spike rate from zero to a substantial value as the threshold is crossed, indicating a subcritical Hopf bifurcation. The system also displays bistability within a specific input power range.
Furthermore, the study explicitly demonstrated the neuron’s frequency preference. The device shows resonant spiking behavior, with output amplitude significantly enhanced when the input frequency matches the neuron’s natural spiking rate. The resonance peak broadens with higher input amplitudes, suggesting that sufficiently strong off-resonance signals can still trigger spiking.
Crucially, the researchers demonstrated the cascadability of these neurons, a critical requirement for building neural networks. On-resonance signals experience regenerative gain, allowing them to propagate through multiple layers of neurons. Off-resonance signals, representing noise or unwanted signals, diminish as they pass through successive layers. The researchers successfully cascaded ten spiking neurons, showing that the output of one neuron can effectively drive the next, proving the potential for building complex neural networks with these devices.
Leave a Reply