Hardware perceptron and ReLU measurements and results

Recently, in-memory analog computing through memristive crossbar arrays
attracted a lot of attention due to its efficient power consumption,
area, and computing throughput. Using this computing method,
different types of neural networks can be implemented for different
applications. In such neural networks, memristors represent the synapses.
However, in previous work, digital processors have been used to
implement the activation functions or neurons. Implementing neurons using
analog-based hardware further improves the power consumption, area, and
throughput by removing unnecessary data conversions and
communication.   In this study, we designed a ReLU
activation function and built a fully hardware-based two-layer fully
connected perceptron using memristive arrays, and verified the
operation by classifying downsampled MNIST images. We measured
the DC and AC characteristics of our designed ReLU, forming, set, and
reset behavior of the memristive arrays, and the
perceptron behavior during training and inference in the
classification task. We also studied the non-idealities related to both
the ReLU design and memristors which is significantly critical in future
integrated designs. Moreover, the downsampled 8*8 MNIST images that we
generated from the original MNIST dataset are included in the
data which can be used in future studies with the limited size of the
network.