Temporal Data Analysis Using Reservoir Computing and Dynamic Memristors

Abstract

Temporal data analysis including classification and forecasting is essential in a range of fields from finance to engineering. While static data are largely independent of each other, temporal data have a considerable correlation between the samples, which is important for temporal data analysis. Neural networks thus offer a more general and flexible approach since they do not depend on parameters of specific tasks but are driven only by the data. In particular, recurrent neural networks have gathered much attention since the temporal information captured by the recurrent connections improves the prediction performance. Recently, reservoir computing (RC), which evolves from recurrent neural networks, has been extensively studied for temporal data analysis as it can offer efficient temporal processing of recurrent neural networks with a low training cost. This dissertation presents a hardware implementation of the RC system using an emerging device - memristor, followed by a theoretical study on hierarchical architectures of the RC system. A RC hardware system based on dynamic tungsten oxide (WOx) memristors is first demonstrated. The internal short-term memory effects of the WOx memristors allow the memristor-based reservoir to nonlinearly map temporal inputs into reservoir states, where the projected features can be readily processed by a simple linear readout function. We use the system to experimentally demonstrate two standard benchmarking tasks: isolated spoken digit recognition with partial inputs and chaotic system forecasting. High classification accuracy of 99.2% is obtained for spoken digit recognition and autonomous chaotic time series forecasting has been demonstrated over the long term. We then investigate the influence of the hierarchical reservoir structure on the properties of the reservoir and the performance of the RC system. Analogous to deep neural networks, stacking sub-reservoirs in series is an efficient way to enhance the nonlinearity of data transformation to high-dimensional space and expand the diversity of temporal information captured by the reservoir. These deep reservoir systems offer better performance when compared to simply increasing the size of the reservoir or the number of sub-reservoirs. Low-frequency components are mainly captured by the sub-reservoirs in the later stages of the deep reservoir structure, similar to observations that more abstract information can be extracted by layers in the late stage of deep neural networks. When the total size of the reservoir is fixed, the tradeoff between the number of sub-reservoirs and the size of each sub-reservoir needs to be carefully considered, due to the degraded ability of the individual sub-reservoirs at small sizes. Improved performance of the deep reservoir structure alleviates the difficulty of implementing the RC system on hardware systems. Beyond temporal data classification and prediction, one of the interesting applications of temporal data analysis is inferring the neural connectivity patterns from the high-dimensional neural activity recording data. By computing the temporal correlation between the neural spikes, connections between the neurons can be inferred using statistics-based techniques, but it becomes increasingly computationally expensive for large scale neural systems. We propose a second-order memristor-based hardware system using the natively implemented spike-timing-dependent plasticity learning rule for neural connectivity inference. By incorporating biological features such as transmission delay to the neural networks, the proposed concept not only correctly infers the direct connections but also distinguishes direct connections from indirect connections. Effects of additional biophysical properties not considered in the simulation and challenges of experimental memristor implementation will be also discussed.