Parylene-based memristive synapses for hardware neural networks capable of dopamine-modulated STDP learning

Abstract

Nowadays there is a growing interest in wearable and biocompatible computing systems that are safe for the human body. Memristive devices are prospective for such tasks owing to a number of their attractive properties, in particular, the multilevel character of resistive switching, or plasticity, which allows them to emulate synapses in hardware neuromorphic networks (NNs). The use of local learning rules for such NNs, for example, bioinspired spike-timing-dependent plasticity (STDP), has firmly established itself in recent years. In biological systems the basic STDP can be modified in the presence of neuromodulators (e.g. dopamine). This effect is believed to be essential for important biological functions such as reinforcement learning (RL), memory and others. The goal of this work was to demonstrate that such dopamine-like modulated STDP can be used in memristors based on a biocompatible polymer, parylene (poly-p-xylylene, or PPX). We have studied memristors both in the form of single Cu/PPX/ITO devices and in the form of crossbar Cu/PPX/Au structures. It was found that, in addition to stable memristive characteristics suitable for NNs, these devices can also change their conductance by means of bioinspired STDP rules, including dopamine-like modulated STDP window realized by introducing the coefficients for neuron spike amplitudes. The amplitude coefficients from −1 (inhibitory mode) to 1 (excitatory mode) of pre-and post-spikes, reflecting the 'dopamine' concentration, in various combinations allow observing the STDP window not only of the usual shape, but also of the anti-STDP, bell and anti-bell shapes. The obtained results demonstrate that the development of memristors based on PPX provides prospects for hardware realization of bio-inspired spiking NNs with RL ability.

Parylene-based memristive crossbar structures with multilevel resistive switching for neuromorphic computing

“…We have previously shown that single memristors work successfully as a part of spiking networks, for example, for associative learning [26]. In addition, in our previous work, we have also successfully demonstrated that both single and crossbar PPX memristors can change their conductance according to the rules of dopamine-modulated STDP rule, and therefore are an important link in the implementation of hardware reinforcement learning [32]. Here we have successfully demonstrated the training of a formal network based on a new improved type of crossbar array of PPX memristors.…”

“…But for the use of memristors in CMOS-compatible hardware neural networks, single structures are unsuitable, and a crossbar topology is required for integration into computational microelectronics [30]. PPX-based memristors can also be realized in arrays of crossbar structures [31,32]. But a detailed study of such structures was not carried out, therefore the purpose of the work was a comprehensive study of the main memristive characteristics, as well as the development of new methods for creating crossbar arrays of PPX-based memristors.…”

Reliable Memristive Synapses Based on Parylene-MoOx Nanocomposites for Neuromorphic Applications

“…We have previously demonstrated that PPX-based memristors without inclusions can be successfully used to construct NCSs. ,, The enhancement of all main characteristics due to the addition of MoO x should undoubtedly lead to an increase in the neuromorphic potential of the memristors. For the neural network modeling, we selected a representative taskMNIST handwritten digits classificationand employed a typical network architecture: a fully connected two-layer network (784 × 512 × 10; see the Methods section for further details).…”

Principles of analog neuromorphic computing: from components to systems and algorithms

“…Продемонстрированы результаты по исследованию устойчивости обучения СНС к вариабельности характеристик мемристоров как аналоговых элементов, а также к использованию шума в качестве конструктивного фактора при обучении и удержании мемристивных весов импульсной сети [7,11]. Обсуждаются также подходы к реализации локальных правил дофаминоподобного обучения с подкреплением в СНС, которые необходимы для формирования аналога системы «потребностей» интеллектуального агента в процессе его автономного функционирования [12][13][14]. Рассмотрены первые результаты по созданию прототипа мемристивного имплантируемого устройства, нейропротезирующего двигательную активность животного [15,16].…”