Fig. 3: Construction of memristor-based feature maps.
From: Memristor-based feature learning for pattern classification
a A flow chart for feature map construction from waveforms. Firstly, the waveform is input into the drift-diffusion kinetics (DDK) equation (algorithm) or transformed to a series of voltage pulses (hardware). Then, the equation solution or memristor conductance response constitutes the feature map. b Hardware-level feature extraction from whale sounds in WMMS. The sound waveforms of blue, dwarf, fin and humpback whales (upper panel) are sub-sampled into 8 utterances and average pooled (1 × 60). The corresponding feature maps are shown in the lower panel. c The device conductance change of the hardware feature map (blue whale). The fitted I-V curves of alpha and beta parameters are shown by orange solid lines, while the measured curves are depicted as blue scattered dots. The fitting accuracy is represented as sum of the squared residuals (R). d Comparison of our method in terms of latency and energy consumption with MFCC-based hardware.
