How does negative differential resistance relate to neuromorphic computing and sensors?

Negative differential resistance, sometimes called negative dynamic resistance (NDR), occurs when an increase in voltage leads to a decrease in current, and vice versa. It’s observed in devices like tunnel diodes, Gunn diodes, IMPATT diodes, and neon lamps, and can be useful in oscillators and other circuits. It’s also being applied to neuromorphic computing and neuromorphic photonic sensory neurons.

Figure 1. Comparison of current-controlled (left) and voltage-controlled (right) NDR. (Image: Wikibooks)

NDR is not negative resistance. The resistance experienced in NDR is positive, and it’s nonlinear. There are two types of NDR, current-controlled and voltage-controlled (Figure 1).

In current-controlled NDR (CC-NDR), also called S-shaped NDR, the voltage is a single-valued function of the current, but the current is a multivalued function of the voltage. At lower voltages, the voltage and current increase together. Once a threshold has been reached, the current starts decreasing as the voltage continues to rise. Current-controlled NDR is found in thyristors like SCRs, unijunction transistors (UJTs), and neon lamps.
In voltage-controlled NDR (VC-NDR), also called N-shaped NDR, the current is a single-valued function of the voltage, but the voltage is a multivalued function of the current. In this case, as the voltage increases, the current increases until a threshold is reached, when the current begins decreasing as the voltage continues to increase. Voltage-controlled NDR is found in tunnel diodes, Gunn diodes, and similar devices.

Expanding applications

Common applications for NDR include high-frequency oscillators and amplifiers in microwave systems, memory devices, and frequency multipliers. Recently, it’s been used to implement neuromorphic computing (NC) and sensor systems. NC is a bio-inspired technology that mimics the neural network in the brain.

Compared with conventional digital computers, NC is highly energy efficient and is being explored for use in artificial intelligence (AI) and machine learning (ML) applications. For example, a Pearson-Anson oscillator employs NDR and exhibits similar behavior to biological neurons, converting DC signals into high-frequency spiking outputs (Figure 2).

Figure 2. Pearson-Anson oscillator using a neon lamp as the NDR element. (Image: Wikipedia)

The use of NC processing and sensing eliminates the need for analog-to-digital conversion. Sensors that produce spike outputs are highly compatible with NC. In sensors, NDR devices can be designed to be sensitive to specific stimuli, like certain gases or wavelengths of light, with the NDR characteristics changing in response to the environment.

Neuromorphic NDR sensors

An artificial sensory oscillator neuron based on VC-NDR has been developed that’s capable of detecting and processing optical information. It consists of a III-V semiconductor micropillar quantum resonant tunnelling diode photodetector (μRTD-PD) with GaAs photosensitive absorption layers.

Figure 3 illustrates the theory, structure, and operation of the μRTD-PD:

Figure 3. The biological sensor oscillator neuron and its optical semiconductor counterpart. (Image: Scientific Reports)

An example of a biological optical sensory oscillatory neuron system found in the human visual system that senses and converts optical signals into electrical oscillations, and was the inspiration for the μRTD-PD.
The μRTD-PD sensory oscillator neuron device structure showing the dimensions of the intrinsic double barrier quantum well (DBQW) region and the surrounding collector and emitter n-type GaAs light absorption layers.
Lumped electrical circuit of the μRTD-PD device. L_s is the connecting circuitry equivalent inductance. R_s is the series resistance and other circuitry resistance. C_RTD is the equivalent device and circuitry capacitance. The f(V, P_in) term is the light and voltage-controlled current source that mimics the RTD current-voltage (I-V) characteristic and depends on the bias voltage and incident light intensity. I_ph is the photocurrent, and I_n is the associated noise of the system.
I-V characteristics of the μRTD-PD oscillatory neuron under dark and illumination conditions. Under dark conditions (black dashed trace), only positive differential resistance regions are present. Under illumination conditions (purple solid line), an N-shaped I-V NDR region occurs.

The μRTD-PD oscillatory neurons are compatible with mature III-V semiconductor processes. They can operate over a wide range of wavelengths and can be readily integrated into industrial sensing, LIDAR, and other optoelectronic applications.

Summary

The two types of NDR, CC-NDR and VC-NDR, are widely used in oscillators, memory devices, and other applications. Currently, both types of NDR are being explored for use in energy-efficient neuromorphic computing and sensing applications to support AI and ML systems. NDR sensors are versatile and can be designed to be sensitive to specific stimuli, such as various gases and wavelengths of light.