2016/06/27 by Christopher H. Bennett, Bennett, Christopher H., Selina La Barbera +7 · 1 voice
Engineering · Computer Science · #Advanced Memory and Neural Computing #Ferroelectric and Negative Capacitance Devices #Machine Learning and ELM
paper · pdf · doi:10.48550/arxiv.1606.08366
Memristive nanodevices offer new frontiers for computing systems that unite\narithmetic and memory operations on-chip. Here, we explore the integration of\nelectrochemical metallization cell (ECM) nanodevices with tunable filamentary\nswitching in nanoscale learning systems. Such devices offer a natural\ntransition between short-term plasticity (STP) and long-term plasticity (LTP).\nIn this work, we show that this property can be exploited to efficiently solve\nnoisy classification tasks. A single crossbar learning scheme is first\nintroduced and evaluated. Perfect classification is possible only for simple\ninput patterns, within critical timing parameters, and when device variability\nis weak. To overcome these limitations, a dual-crossbar learning system partly\ninspired by the extreme learning machine (ELM) approach is then introduced.\nThis approach outperforms a conventional ELM-inspired system when the first\nlayer is imprinted before training and testing, and especially so when\nvariability in device timing evolution is considered: variability is therefore\ntransformed from an issue to a feature. In attempting to classify the MNIST\ndatabase under the same conditions, conventional ELM obtains 84%\nclassification, the imprinted, uniform device system obtains 88%\nclassification, and the imprinted, variable device system reaches 92%\nclassification. We discuss benefits and drawbacks of both systems in terms of\nenergy, complexity, area imprint, and speed. All these results highlight that\ntuning and exploiting intrinsic device timing parameters may be of central\ninterest to future bio-inspired approximate computing systems.\n