Published in Nanoscale Horizons journal, under the title “Fully transparent, flexible and waterproof synapses with pattern recognition in organic environments”, the e-synapses’ reported synaptic behaviors included paired-pulse facilitation (PPF), long-term potentiation/depression (LTP/LTD), and learning–forgetting–relearning. Made of a layer of poly(3,4-ethylenedioxythiophene)/poly-styrene sulfonate (PEDOT/PSS) sandwiched between indium tin oxide (ITO) electrodes, the artificial synaptic device exhibited an optical transmittance of 87.5% in the visible light range, it operated reliably even when bent to a 5mm radius and was proven to withstand water and five types of common organic solvents for over 12 hours, functioning over 6000 spikes without noticeable degradation.
Using the transparent, flexible, and biocompatible e-synapses, the researchers then constructed a three-layer neural network system consisting of an input layer, a hidden layer, and an output layer connected with 256×128 and 128×10 e-synapses. The authors used resampled 16×16 pixel images of handwritten digits from the MNIST database (to match the 256 input neurons) and trained the multilayer perceptron network with 10,000 initial sample images based on back-propagation, keeping 500 images to verify the recognition effect. The article report a 92.4% recognition accuracy for the handwritten digits, promising lightweight, transparent, flexible and biocompatible neuromorphic computing systems.
Fudan University – www.fudan.edu