ABSTRACT
MLP-Mixer based on multilayer perceptrons (MLPs) is a novel architecture of a neuromorphic computing system (NCS) introduced for image classification tasks without convolutional layers. Its software realization demonstrates high classification accuracy, although the number of trainable weights is relatively low. One more promising way of improving the NCS performance, especially in terms of power consumption, is its hardware realization using memristors. Therefore, in this work, we proposed an NCS with an adapted MLP-Mixer architecture and memristive weights. For this purpose, we used a passive crossbar array of (Co-Fe-B)x(LiNbO3)100-x memristors. Firstly, we studied the characteristics of such memristors, including their minimal resistive switching time, which was extrapolated to be in the picosecond range. Secondly, we created a fully hardware NCS with memristive weights that are capable of classification of simple 4-bit vectors. The system was shown to be robust to noise introduction in the input patterns. Finally, we used experimental memristive characteristics to simulate an adapted MLP-Mixer architecture that demonstrated a classification accuracy of (94.7 ± 0.3)% on the Modified National Institute of Standards and Technology (MNIST) dataset. The obtained results are the first steps toward the realization of memristive NCS with a promising MLP-Mixer architecture.
ABSTRACT
Convolutional neural networks (CNNs) have been widely used in image recognition and processing tasks. Memristor-based CNNs accumulate the advantages of emerging memristive devices, such as nanometer critical dimensions, low power consumption, and functional similarity to biological synapses. Most studies on memristor-based CNNs use either software models of memristors for simulation analysis or full hardware CNN realization. Here, we propose a hybrid CNN, consisting of a hardware fixed pre-trained and explainable feature extractor and a trainable software classifier. The hardware part was realized on passive crossbar arrays of memristors based on nanocomposite (Co-Fe-B)x(LiNbO3)100-x structures. The constructed 2-kernel CNN was able to classify the binarized Fashion-MNIST dataset with ~ 84% accuracy. The performance of the hybrid CNN is comparable to the other reported memristor-based systems, while the number of trainable parameters for the hybrid CNN is substantially lower. Moreover, the hybrid CNN is robust to the variations in the memristive characteristics: dispersion of 20% leads to only a 3% accuracy decrease. The obtained results pave the way for the efficient and reliable realization of neural networks based on partially unreliable analog elements.
ABSTRACT
In this paper, the resistive switching and neuromorphic behaviour of memristive devices based on parylene, a polymer both low-cost and safe for the human body, is comprehensively studied. The Metal/Parylene/ITO sandwich structures were prepared by means of the standard gas phase surface polymerization method with different top active metal electrodes (Ag, Al, Cu or Ti of ~500 nm thickness). These organic memristive devices exhibit excellent performance: low switching voltage (down to 1 V), large OFF/ON resistance ratio (up to 104), retention (≥104 s) and high multilevel resistance switching (at least 16 stable resistive states in the case of Cu electrodes). We have experimentally shown that parylene-based memristive elements can be trained by a biologically inspired spike-timing-dependent plasticity (STDP) mechanism. The obtained results have been used to implement a simple neuromorphic network model of classical conditioning. The described advantages allow considering parylene-based organic memristors as prospective devices for hardware realization of spiking artificial neuron networks capable of supervised and unsupervised learning and suitable for biomedical applications.
