Energy-Efficient Pattern Recognition Hardware With Elementary Cellular Automata

Polaritonic Neuromorphic Computing Outperforms Linear Classifiers

“…Even though a direct comparison between different physical architectures is difficult, our system is able to obtain higher accuracy than previously reported RC approaches while working with lower-resolution images. In Table , different strategies are compared, including the realization of terahertz deep NN with metallic masks (81% MNIST accuracy), a convolutional approach in which the hopping conduction in a silicon contact is controlled by artificial evolution (96% MNIST accuracy), and a cellular automata approach to RC (ReCA) implemented on an FPGA (98% MNIST accuracy). While state-of-the-art software realizations of NNs achieve accuracies on the MNIST data set higher than 99.8%, the potentiality of photonic NN can be most appreciated in reducing the energy consumption or accelerating specific operations in complex problems of image classification. − The picosecond dynamics and high accuracy of polariton lattices therefore seem to be suited for image recognition when fast operational speed, noisy information, or offline devices are involved.…”

“… a The sizes of the testing and training sets as well as the resolution of the input digits are reported together with the success rate. The speed is an estimated maximum operational frequency for digit classification based only on the fundamental physical limits of the system. b The hopping current experiment in ref is not a full hardware implementation but rather a software simulation of thousands of identical dopant systems. c The ReCA system in ref was trained with an extended MNIST data set including distortions, which improved the recognition rate. …”

“… c The ReCA system in ref was trained with an extended MNIST data set including distortions, which improved the recognition rate. …”

The Dynamical Landscape of Reservoir Computing with Elementary Cellular Automata

“…CA can be considered to be a special case of (neural) networks, where connectivity is ordered such that every node is connected to exactly the same number of neighbours, which in turn are uniformly connected to each-other. One advantage of CA is that they can be implemented in hardware, such as in Field Programmable Gate Arrays (FPGA) (Morán et al, 2019), which results in energy efficient computing substrates as well as fast execution due to the hardware supporting CA implementation natively and CA being fully parallel.…”

Cellular automata can classify data by inducing trajectory phase coexistence

“…Cellular automata have been used to do computations such as classification, either directly [7,8], as a reservoir that enacts a nonlinear transformation prior to additional calculation [9][10][11], or in combination with a neural network [12]. Cellular automata do not currently perform as well at most machine-learning tasks as deep neural networks, but continue to attract attention because discrete automata specified by integers consume less memory and power than do real-valued neural nets, a property useful for low-power mobile devices [13].…”