Memristive Pixel-CNN Loop Generate for CNN Generalisations

Noisein test images can significantly reduce the inference accuracy of Convolution Neural Networks (CNNs). The learning alogrithm that optimise the CNN for inference often use training images that represent a limited number of variations of objects. Any variations from the training set makes it harder to recognise test images, thereby indicating lose of generalisation. We propose to use Gated Pixel Convolution Neural Network (PixelCNN) for generating training images to readjust the weights of pretrained CNN network. This way, the readjustment of weights using generated images, helps to improve the generalisation ability of the CNNs. Through this approach the CNN learning becomes continuous even with a limited training data and can be limited by the amount of generalisation and robustness to variability required for a given task. The proposed memristive loop generate architecture was designed using 22 nm CMOS technology and Knowm Meta-Stable Switch (MSS) memristor model (
$WO_{x}$
device parameters) with an on-chip area of
$\text{1.54}\;\text{mm}^{2}$
and reduced power consumption.