Spike timing-dependent plasticity using memristors and nano-crystalline silicon TFT memories

Interest in the possibility of using memristive devices as synapses in artificial neural circuits was sparked by the demonstration of TiO₂ resistive switches by HP Labs in 2008 [1]. A great deal of the resulting research has been centered on implementing spike timing-dependent plasticity (STDP), which is a synaptic learning mechanism based on timing differences between action potentials [2,3]. However, time scales of biological inter-spike intervals (ISIs) are on the order of tens of milliseconds, much longer than the typical electronic phenomena. This makes STDP a difficult scheme to implement efficiently using electronics. Proposed solutions have involved pulse width or height modulation [4-6] or pulse shaping [7] and would require somewhat extensive circuitry for each neuron. Additionally, the reports do not explain the learning characteristics beyond pair-based trials. Experiments on biological synapses indicate a much more complex reality, in that the exact mechanisms of synaptic learning cannot be explained by pair-based STDP alone [8,9]. Specifically, asymmetric temporal integration of the synaptic weight changes has been demonstrated in spike triplet and quadruplet, as well as frequency-dependent experiments [10-12]. Progress continues on developing models that explain the observed effects more thoroughly [13,14].