Recently, several powerful machines dedicated to solving combinatorial optimization problems through the Ising-model formulation have appeared. The trigger for the paradigm shift to a specialized machine for solving optimization problems was the D-Wave machine, which implements quantum annealing. Quantum annealing employs quantum fluctuations to find an optimal solution to an optimization problem with discrete variables. In particular, we input the optimization problem in the form of the Ising Hamiltonian, which is a specialized form of the quadratic unconstrained binary optimization problem. However, when we employ quantum annealing for a practical optimization problem, there are several issues. One typical issue is absence of the detailed form of the cost function, which characterizes the optimization problem. To input problems into specialized machines for solving an optimization problem, it is necessary to determine the unknown parameters within the Ising model. We propose a method to estimate the unknown parameters in the Ising Hamiltonian using compressed sensing. Furthermore, we analyze the theoretical limitations of our proposed method by employing the replica method, which is a sophisticated tool in statistical mechanics.
This paper reviews applications of stochastic computing in brainware LSI (BLSI) for visual information processing. Stochastic computing exploits random bit streams, realizing the area-efficient hardware of complicated functions, such as multiplication and tanh functions in comparison with binary computation. Using stochastic computing, we implement the hardware of several physiological models of the primary visual cortex of brains, where these models require such the complicated functions. Our vision BLSIs are implemented using Taiwan Semiconductor Manufacturing Company (TSMC) 65 nm CMOS process and discussed with traditional fixed-point implementations in terms of hardware performance and computation accuracy. In addition, an analog-to-stochastic converter is designed using CMOS and magnetic tunnel junctions that exhibit probabilistic switching behaviors for area/energy-efficient signal conversions to stochastic bit streams.
This paper presents a 32-channel compressive gammachirp filterbank chip based on hybrid stochastic/binary computation for area/power-efficient auditory signal processing. The gammachirp filter well expresses the performance of human auditory peripheral mechanism and can be used for hearing assisting devices and noise robust speech recognition systems. The stochastic gammachirp filters are designed using cascaded digital IIR filters, leading to area-efficient hardware thanks to a simple logic-gate implementation of multiplication. However, the signal variability due to random number sequences used in stochastic computation induces unwanted frequency components at each IIR filter, causing large noise signals at the output of the gammachirp filters. To reduce the noise signals, a fixed random-number-generation (FRNG) technique is introduced that provides the same random number sequence at every operation as opposed to different random number sequences used in a conventional stochastic filter. The FRNG technique mitigates the noise signals and hence increases the filter gains with short lengths of stochastic bit streams. In addition, gain-compression characteristics depending on input acoustic pressures known as human auditory effects are naturally realized by changing the lengths of the stochastic bit streams. The proposed filterbank chip is fabricated using Taiwan Semiconductor Manufacturing Company (TSMC) 65 nm CMOS process that achieves 715-2,585 µW with the chip area of 3.2 mm2, leading to the best power-area product per channel in comparison with conventional analog auditory filterbanks.
In this paper, a novel spike-train generator suitable for quantum-dot cellular automaton (QCA) implementation is proposed. As an analysis tool of the proposed generator, a novel spike phase map is derived. Then, using the map, it is shown that the proposed generator can generate spike-trains with various spike patterns in terms of period, density, and correlation. Furthermore, a stochastic algorithm for parameter tuning for the proposed generator is proposed. It is shown that the parameter tuning algorithm enables the proposed generator to generate spike-trains suitable for ultra wide band impulse radio (UWB-IR) communication, ranging, and positioning systems. A QCA layout of the generator obtained by the parameter tuning algorithm is designed and its operation is verified by a QCA simulator.
We propose “QER”, a novel regularization strategy for hardware-aware neural network training. Although quantized neural networks reduce computation power and resource consumption, it also degrades the accuracy due to quantization errors of the numerical representation, which are defined as differences between original numbers and quantized numbers. The QER solves such the problem by appending an additional regularization term based on quantization errors of weights to the loss function. The regularization term forces the quantization errors of weights to be reduced as well as the original loss. We evaluate our method by using MNIST on a simple neural network model. The evaluation results show that the proposed approach achieves higher accuracy than the standard training approach with quantized forward propagation.
Neural networks architectures will be soon required to be deployed on IoT-edge systems, which will demand small and low-power circuits and application-dependent customizability. Focusing on the power-efficiency and internal structures of 3D memristive devices, this paper proposes neural networks architectures that are implementable in those devices, together with an existing transfer learning technique. Our architectural model is stochastic and device-conscious in handling noise/variation-vulnerability and suppressing inter-layer connections. Through experiments, we quantitatively explored appropriate structures under device constraints and demonstrated comparable performance as conventional work holding complex connections. Also, we revealed an important challenge on device technologies for further performance improvement.
This paper presents an efficient iterative method to solve standard and generalized algebraic Riccati equations for RLC networks. The proposed method generates a low-rank solution of the Riccati equation for a positive real balanced truncation. Linear passive RLC networks are index-1 or -2 systems;the proposed method generates a low-rank solution of the standard algebraic Riccati equation for an index-1 system and that of a generalized Riccati equation for an index-2 system. To generate accurate reduced-order models at low and high frequencies, the parameters of the iterative method for solving the Riccati equations are investigated. In general, the balanced truncation accuracy at low frequencies is not necessarily satisfactory, compared to that of (Krylov) projection-based model-order reduction. The accuracy of a reduced-order model at low frequencies can be improved by considering a constant shift parameter.
Tracking optima in dynamic problems is achieved by a multi-population optimizer based on piecewise-rotational chaotic system(OPRC) using memory update within a tolerance. Tracking optima is a difficult task for multi-population-based optimizers because of two issues, called outdated memory and divergent loss. To solve the outdated memory issue, a simple procedure named memory update within a tolerance is proposed. The proposed procedure is applied to our previous proposed optimizer OPRC, and its outstanding tracking performance is observed. This result shows OPRC can solve the divergent loss issue without any modification of its searching dynamics. The tracking mechanism of OPRC is also considered, and it is uncovered that the searching behavior given by the folding dynamics of the chaotic system contributes to catch the shifting optima.