//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>
Compact, low–latency, and low–power computer systems are required for real–world sensory–processing applications. Hybrid memristive CMOS neuromorphic architectures, with their in–memory event–driven computing capabilities, present an appropriate hardware substrate for such tasks.
To demonstrate the full potential of such systems and drawing inspiration from the barn owl’s neuroanatomy, CEA–Leti has developed an event–driven, object–localization system that couples state–of–the–art piezoelectric, ultrasound transducer sensors with a neuromorphic computational map based on resistive random–access memory (RRAM).
CEA–Leti built and tested this object tracking system with the help of researchers from CEA–List, the University of Zurich, the University of Tours, and the University of Udine.
The researchers conducted measurements findings from a system built out of RRAM–based coincidence detectors, delay–line circuits, and a fully customized ultrasonic sensor. This experimental data has been used to calibrate the system–level models. These simulations have then been used to determine the object localization model’s angular resolution and energy efficiency. Presented in a paper published recently in Nature Communications, the research team describes the development of an auditory–processing system that increases energy efficiency by up to five orders of magnitude compared with conventional localization systems based on microcontrollers.
“Our proposed solution represents a first step in demonstrating the concept of a biologically inspired system to improve efficiency in computation,” said Elisa Vianello, senior scientist and edge AI program coordinator and senior author of the paper. “It paves the way toward more complex systems that perform even more sophisticated tasks to solve real–world problems by combining information extracted from different sensors. We envision that such an approach to conceive a bio–inspired system will be key to build the next generation of edge AI devices, in which information is processed locally and with minimal resources. In particular, we believe that small animals and insects are a great source of inspiration for an efficient combination of sensory information processing and computation. Thanks to the latest advancements in technology, we can couple innovative sensors with advanced RRAM–based computation to build ultra–low–power systems.”
Bio–inspired analog RRAM–based circuit
Two essential ideas underpin biological signal processing: event–driven sensing and in–memory analog processing.
“The goal is, as always, to get the best power efficiency for the level of performance needed by a specific application,” Vianello said. “Further improvements in energy efficiency are certainly possible with our system. For example, one could optimize our design and implement it in a more advanced technological node or with a specific low–power technology such as FD–SOI for the same level of performance. Concerning accuracy, our limiting factor is SNR. We have a clear performance/consumption tradeoff with the amplitude of the emitted pulse or the number of TX membranes, but technological advancement resulting in increased piezoelectric micromachined ultrasonic transducer [pMUT] sensitivity would also help improve the SNR for no extra power consumption. The use of pulses with good autocorrelation properties would be an interesting development in that sense if the matched filtering could be done with a small overhead.”
The team leveraged CEA–Leti’s successes in building pMUTs and its developments in RRAM–based spiking neural networks. The initial difficulty for the researchers was to create a pre–processing pipeline that pulls critical information from pMUTs, which encode information using brief events or spikes. This temporal encoding of the signal saves energy over standard continuous analog or digital data because only relevant data is handled.
PMUTs are becoming one of the most demanding ultrasonic systems due to their ability to create and detect ultrasound signals at the microscale in a highly efficient and well–controlled manner. The high–yield MEMS production technique, combined with thin–film piezoelectric materials (AlN, AlScN, PZT, etc.), enhances PMUT systems. Furthermore, the ability to install thin–film piezoelectric materials in a CMOS–compatible manner opens the door to innovative, extremely small systems that use the same substrate for the sensor and the conditioning electronics.
With this scenario, PMUT transducers are pushing the applicability of ultrasound as a physical magnitude in a variety of systems where size, power, sensitivity, and cost are important. These include intravascular medical imaging, biometric identification, gesture recognition, rangefinders, proximity sensors, acoustic wireless communication systems, acoustophoresis, photoacoustic systems, and so on.
According to Vianello, pMUT devices are mature for industrialization. “One of the main restrictions to the development of pMUT devices is the competition of bulk PZT transducer and cMUT MEMs transducers. Bulk PZT transducers are easy to prototype and relatively cheap for low–volume production. cMUT MEMS transducers are more appropriate for biomedical applications due to their higher bandwidth and higher output pressure. One of the physical limitations of pMUT is the relatively low Q factor that results in transient regime that is detrimental to the spatial resolution and may impede short–distance measurements. Industrially matured piezoelectric materials for pMUT are PZT and AlN. PZT is more appropriate for actuating and AlN for sensing. For this application, we need both actuation and sensing, and our approach would have been valid with either of these materials. Yet we choose AlN because the four–electrode–pair scheme, which is not possible with PZT material, partially balances the relatively low output pressure per volt. Moreover, output pressure may be easily increased by the use of higher actuation voltage, at the price of higher consumption.”
Another difficulty was developing and building an analog circuit based on biologically inspired RRAM to analyze extracted events and estimate an object’s location. RRAM is a non–volatile technology that suits the asynchronous nature of events in the team’s proposed system, resulting in negligible power usage while the system is idle.
RRAM stores information in its non–volatile conductive state. The primary operational assumption of this technology is that altering the atomic state via precise programming operations controls the conductance of the device.
The researchers used an oxide–based RRAM with a 5–nm hafnium–dioxide layer sandwiched between top and bottom electrodes made of titanium and titanium nitride. By applying current/voltage waveforms that construct or break a conductive filament made up of oxygen vacancies between the electrodes, the conductivity of an RRAM device may be changed. They co–integrated these devices in a standard 130–nm CMOS process to build a reconfigurable neuromorphic circuit that included coincidence detectors and delay–line circuits (Figure 1). The non–volatile and analog nature of these devices perfectly match the event–driven nature of the neuromorphic circuits, resulting in low power consumption.
The circuit has an instant on/off feature: It begins operating immediately after being turned on, allowing the power supply to be entirely shut off as soon as the circuit is idle. Figure 1 displays the basic building block of the proposed circuit. It is composed of N parallel one–resistor–one–transistor (1T1R) structures that contain synaptic weights and is used to extract a weighted current that is then injected into a common differential pair integrator (DPI) synapse and subsequently into a leaky integrate–and–fire (LIF) neuron.
The input spikes are applied to the gates of the 1T1R structures as trains of voltage pulse with pulse lengths in the range of hundreds of nanoseconds. RRAM may be set into a high–conductance state (HCS) and reset into a low–conductance state (LCS) by providing an external positive voltage reference on Vtop and grounding Vbottom (LCS). The mean value of the HCS may be controlled by limiting the set programming (compliance) current (ICC) through the gate–source voltage of the series transistor. In the circuit, RRAMs perform two functions: They route and weigh input pulses.
“The op amp in Figure 1, along with transistors M1, M2, and M3, form the front–end circuit, which reads the current from the RRAM array and injects the current into the DPI synapse,” Vianello said. “The RRAM bottom electrode has a constant DC voltage Vbot applied to it, and the common top electrode is pinned to the voltage Vx by a rail–to–rail operational–amplifier circuit. The op–amp output is connected in negative feedback to its non–inverting input and has the constant DC bias voltage Vtop applied to its inverting input. As a result, the output of the op amp will modulate the gate voltage of transistor M1 such that the current it sources onto the node Vx will maintain its voltage as close as possible to the DC bias Vtop. Whenever an input pulse Vin arrives, a current equal to (Vx − Vbot)Gn will flow out of the bottom electrode. The negative feedback of the op amp will then act to ensure that Vx = Vtop by sourcing an equal current from transistor M1. By connecting the op–amp output to the gate of transistor M2, a current equal to it will therefore also be buffered into the branch composed of transistors M2 and M3 in series. This current is injected into a CMOS differential–pair integrator synapse circuit model, which generates an exponentially decaying waveform from the onset of the pulse with an amplitude proportional to the injected current.”
While traditional processing techniques sample the detected signal continuously and perform calculations to extract useful information, the proposed neuromorphic solution calculates asynchronously when useful information arrives, increasing the system’s energy efficiency by up to five orders of magnitude.
CEA–Leti has made significant developments in pMUT sensors and spiking neural networks based on RRAM technology during the last decade. “Thank the H2020 MeM–Scales project  that partially funded the work,” Vianello said.
The present study demonstrates that combining visual sensors such as DVS cameras with the suggested pMUT–based hearing sensor should be investigated to create future consumer robots.