A reservoir computing system with volatile and non-volatile organic memristors as a promising hardware architecture

• DOI: 10.17816/gc623423
• Publication Date: 2024-05-06T08:05:41Z
• Authors: A. N. Matsukatova; N. V. Prudnikov; V. A. Kulagin; A. D. Trofimov; A. V. Emelyanov
• Keywords: Memristor

In recent years, many scientific groups have been working on hardware implementation of the artificial neural networks to approach computational efficiency their biological counterpart. Memristors may play role synapses in such [1]. Varieties memristive structures and materials already tested different network architectures, but still no memristor is considered ideal for synapse One most significant problems presence inherent stochasticity distinctive all devices, which complicates training Several approaches were proposed partially mitigate this problem, e.g., a reservoir computing system (RCS) [2] spiking (SNN) [3] as well defect engineering characteristics improvement. work, we propose combine RCS with SNN create bio-inspired neuromorphic based two types organic memristors specifically designed advanced characteristics. The consists main parts: readout [2]. layer extracts some representative features from input data due its internal nonlinear dynamics. then uses these classify data. Typically, conventional fully connected used RCS. process occurs only layer, while not trainable. This decrease trainable parameters considerably reduces impact process. use essential. should consist short-term memory, i.e., volatile memristors. way, can each sample individually. Volatile polyaniline-based chosen implementation. They operate within biologically plausible time range, essential aim mimic systems [4]. contrast, long-term non-volatile memristors, because preserve trained synaptic weights. Non-volatile parylene incorporated MoO3 nanoparticles layer. adopts principles brain function, both short- memory are systems. However, traditional commonly RCSs Their requires global weight updates, making them vulnerable stochasticity. SNNs allow local training, using learning rules, makes more effective robust [3]. Consequently, presume that an promising architecture. work software. First, polyaniline- parylene-based devices fabricated tested. Hardware polyaniline demonstrated ability extract characteristic Nanocomposite suitable unique combination high switching speed, stability, low power consumption possibility crossbar Next, layers compared simulation. It was shown adaptive sustainable noise image classification tasks [5].