While there are several ways to implement artificial neural networks on silicon chips, current-mode implementations are often unscalable and wasteful. Depending on the implementation technique, the neural network may be implemented on one chip or distributed over several. The following sections discuss two methods that can be used to implement artificial neurons. In each case, a new technique is introduced. A new technique is used to model the firing time of a single neuron.
In this breakthrough, researchers developed artificial neurons on silicon chips that mimic the function of biological neurons. These chips need only a billionth of the power of a microprocessor to function. If these chips are integrated into bio-electronic devices, they could help repair damaged biocircuits. In addition to repairing diseased biocircuits, these devices may cure many ailments that are the result of damaged or destroyed neurons.
The deterministic LIF neuron in Figure 3D has a refractory period, which depends on the input voltage. In a similar fashion, an LIF neuron can be programmed to respond to changing input pulse widths by tuning the interval between the pulses. In this way, the firing rate can be controlled by tuning the width of the input pulses. In addition to LIF neuron behavior, Figure 3F shows the output spikes produced by an excitatory input current pulse. The smaller the input pulse interval, the smaller the probability of firing.
A VCM-based artificial neuron implements the principles of a classical LIF neuron. The VO2 device is coupled to a tunable resistor. In this case, a memristor is used as the threshold. The device has a threshold voltage that can be tuned. The frequency of the output spikes is also tuned by tuning the threshold voltage. Its maximum frequency can reach 100 kHz.
Although ANNs are highly accurate, stochastic neurons are likely to play a critical role in embedded systems. Because of their low-energy consumption per spike, stochastic neurons need to be optimized for lower power. They should also have low energy consumption per spike in order to maintain functionality for longer. Further analysis is needed to improve the accuracy of these neurons. Additionally, it is necessary to develop dedicated investigations to understand the device’s dynamics.
One method to build artificial neurons is based on the principles of bio-based computing. The artificial neurons are made up of silicon microcircuits that mimic the structures and functions of live neurons. In addition, the USC team has already built circuits modeled after 100 neurons. And in the near future, the process will be scalable to a million neurons. The ultimate goal of this process is to create a system that can be used to simulate the human brain.
In biophysical neuron models, the spiking threshold module is responsible for setting the voltage at which a neuron spikes. Once the voltage reaches this threshold, the LGP leaks, releasing the potential and returning to a resting state. Similarly, the spiking threshold module controls the maximum firing rate of the neuron. Once the firing rate reaches a threshold, the spiking threshold module activates a positive feedback module. This reduces the time it takes for the inverters to switch polarity and improves the power consumption.