Memristors are promising candidates for synapse emulation in brain-inspired neuromorphic computing systems. The main obstacle to their usage in such systems is high variability of memristive characteristics and its severe negative effect on the neural network training. This paper addresses the issue from two points of view on the example of the parylene-based memristors: (i) the methods of the memristor internal stochasticity decrease and (ii) the methods of the memristive neural network architecture simplification. The introduction of an optimal Ag nanoparticle concentration (3 vol.%–6 vol.%) to the memristive structure leads to a statistically significant decrease in the switching voltage variation and endurance increase. Moreover, it is shown that post-fabrication annealing improves memristive characteristics, e.g., resistive switching window increases by an order of magnitude and exceeds 106, the switching voltage variation decreases by a factor of 2 (down to 7% for the set and 17% for the reset voltage), and thermostability is improved. Additional transmission electron microscopy and impedance spectroscopy analysis allowed establishing a multifilamentary resistive switching mechanism for nanocomposite parylene-based memristors. The simulation of the formal neural network based on these memristors demonstrates high classification accuracy with low variation for an important biomedical task, heart disease prediction, after careful feature selection and network architecture simplification. Future prospects of the controlled incorporation of the nanocomposite parylene-based memristors in neural networks are brightened by their scaling possibility in crossbar geometry.
Parylene is a widely used polymer possessing advantages such as simple and cheap production, possibility of fabrication on flexible substrates, transparency, and safety for the human body. Moreover, parylene can be used as an active layer of memristors—circuit design elements that are promising for the implementation of hardware neuromorphic systems. Recent studies show that memristors are not merely memory but also highly dynamical systems that can encode timing information. Here, a study of the switching kinetics and the timing second-order effects in memristors based on pristine and nanocomposite (with embedded silver nanoparticles) parylene is presented. The strong decrease in the resistive switching time and increase in the amplitude of the resistive state change after preliminary heating pulses are revealed. These effects are explained by the local heating of the parylene matrix by electric pulses, and the given explanation is supported by the numerical electrothermal model. Spike-timing-dependent plasticity with symmetrical nonoverlapping spikes is demonstrated. The obtained results indicate a possibility of the utilization of second-order effects in the development of the neuromorphic systems.
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.
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.
14 November 2023 in ACS Applied Materials & Interfaces
Reliable parylene–PbTe memristors controlled via electrical and optical stimuli replicate key synaptic functions and are applicable in neuromorphic computing systems.