Fig. 1.

Stochastic dot-product circuit and its applications in neurocomputing. a Circuit schematics for the design with current-mode sensing with crosspoint device implementation based on (b) memristors and (c) floating-gate memories. The equations in figure corresponds to memristor implementation, while their modified version for floating-gate design are described in text. d An example of the considered differential-pair Boltzmann machine implementation. e The implementation of generalized Hopfield neural network. The blue background highlights the baseline implementation. The yellow, green, and red backgrounds highlight additional circuitry for the proposed “stochastic”, “adjustable”, and “chaotic” approaches, respectively. The gray shaded circles show synaptic weights which are typically set to zero. Labels “Σ”/ “x” inside amplifier symbols denote summation / scaling. For clarity, panel a does not show bias currents, panels (a) and (e) show single-ended network, while panel (d) shows (a) small two-input, two-hidden neurons fragment of the considered network