Photoelectroactive artificial synapse and its application to biosignal pattern recognition

In recent years, optoelectronic artificial synapses have garnered a great deal of research attention owing to their multifunctionality to process optical input signals or to update their weights optically. However, for most optoelectronic synapses, the use of optical stimuli is restricted to an excitatory spike pulse, which majorly limits their application to hardware neural networks. Here, we report a unique weight-update operation in a photoelectroactive synapse; the synaptic weight can be both potentiated and depressed using “optical spikes.” This unique bidirectional operation originates from the ionization and neutralization of inherent defects in hexagonal-boron nitride by co-stimuli consisting of optical and electrical spikes. The proposed synapse device exhibits (i) outstanding analog memory characteristics, such as high accessibility (cycle-to-cycle variation of <1%) and long retention (>21 days), and (ii) excellent synaptic dynamics, such as a high dynamic range (>384) and modest asymmetricity (<3.9). Such remarkable characteristics enable a maximum accuracy of 96.1% to be achieved during the training and inference simulation for human electrocardiogram patterns.

Toward Software-Equivalent Accuracy on Transformer-Based Deep Neural Networks With Analog Memory Devices

Recent advances in deep learning have been driven by ever-increasing model sizes, with networks growing to millions or even billions of parameters. Such enormous models call for fast and energy-efficient hardware accelerators. We study the potential of Analog AI accelerators based on Non-Volatile Memory, in particular Phase Change Memory (PCM), for software-equivalent accurate inference of natural language processing applications. We demonstrate a path to software-equivalent accuracy for the GLUE benchmark on BERT (Bidirectional Encoder Representations from Transformers), by combining noise-aware training to combat inherent PCM drift and noise sources, together with reduced-precision digital attention-block computation down to INT6.

A Versatile, Voltage-Pulse Based Read and Programming Circuit for Multi-Level RRAM Cells

In this work, we present an integrated read and programming circuit for Resistive Random Access Memory (RRAM) cells. Since there are a lot of different RRAM technologies in research and the process variations of this new memory technology often spread over a wide range of electrical properties, the proposed circuit focuses on versatility in order to be adaptable to different cell properties. The circuit is suitable for both read and programming operations based on voltage pulses of flexible length and height. The implemented read method is based on evaluating the voltage drop over a measurement resistor and can distinguish up to eight different states, which are coded in binary, thereby realizing a digitization of the analog memory value. The circuit was fabricated in the 130 nm CMOS process line of IHP. The simulations were done using a physics-based, multi-level RRAM model. The measurement results prove the functionality of the read circuit and the programming system and demonstrate that the read system can distinguish up to eight different states with an overall resistance ratio of 7.9.

Analog Memristive Characteristics of Square Shaped Lanthanum Oxide Nanoplates Layered Device

Square-shaped or rectangular nanoparticles (NPs) of lanthanum oxide (LaOx) were synthesized and layered by convective self-assembly to demonstrate an analog memristive device in this study. Along with non-volatile analog memory effect, selection diode property could be co-existent without any implementation of heterogeneous multiple stacks with ~1 μm thick LaOx NPs layer. Current–voltage (I–V) behavior of the LaOx NPs resistive switching (RS) device has shown an evolved current level with memristive behavior and additional rectification functionality with threshold voltage. The concurrent memristor and diode type selector characteristics were examined with electrical stimuli or spikes for the duration of 10–50 ms pulse biases. The pulsed spike increased current levels at a read voltage of +0.2 V sequentially along with ±7 V biases, which have emulated neuromorphic operation of long-term potentiation (LTP). This study can open a new application of rare-earth LaOx NPs as a component of neuromorphic synaptic device.

Continuous-Time Programming of Floating-Gate Transistors for Nonvolatile Analog Memory Arrays

Floating-gate (FG) transistors are a primary means of providing nonvolatile digital memory in standard CMOS processes, but they are also key enablers for large-scale programmable analog systems, as well. Such programmable analog systems are often designed for battery-powered and resource-constrained applications, which require the memory cells to program quickly and with low infrastructural overhead. To meet these needs, we present a four-transistor analog floating-gate memory cell that offers both voltage and current outputs and has linear programming characteristics. Furthermore, we present a simple programming circuit that forces the memory cell to converge to targets with 13.0 bit resolution. Finally, we demonstrate how to use the FG memory cell and the programmer circuit in array configurations. We show how to program an array in either a serial or parallel fashion and demonstrate the effectiveness of the array programming with an application of a bandpass filter array.

Quasiperiodic-Chaotic Neural Networks and Short-Term Analog Memory

A model of quasiperiodic-chaotic neural networks is proposed on the basis of chaotic neural networks. A quasiperiodic-chaotic neuron exhibits quasiperiodic dynamics that an original chaotic neuron does not have. Quasiperiodic and chaotic solutions are exclusively isolated in the parameter space. The chaotic domain can be identified by the presence of a folding structure of an invariant closed curve. Using the property that the influence of perturbation is conserved in the quasiperiodic solution, we demonstrate short-term visual memory in which real numbers are acceptable for representing colors. The quasiperiodic solution is sensitive to dynamical noise when images are restored. However, the quasiperiodic synchronization among neurons can reduce the influence of noise. Short-term analog memory using quasiperiodicity is important in that it can directly store analog quantities. The quasiperiodic-chaotic neural networks are shown to work as large-scale analog storage arrays. This type of analog memory has potential applications to analog computation such as deep learning.