Bipolar Morphological Neural Networks: Gate-Efficient Architecture for Computer Vision

The priority of building hardware-oriented neural network models is growing steadily. target goals for their development are the performance and energy efficiency promising hardware-software solutions. Simultaneously, different classes computing architectures computer, optimal will differ. most interesting from a practical point view application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) central processing units (CPUs). We have recently proposed bipolar morphological as model these computer types, computationally intensive parts which use only maximum addition. In this work, we present first time theoretical assessment expressive power consisting BM neurons show that it corresponds to classical multilayer perceptron. addition, summarize current results on in typical tasks technical vision: image classification semantic segmentation. consider simple LeNet-5-like networks, well deeper ResNet UNet architectures. networks demonstrate accuracy allows use, with significantly higher terms transistor budget two (ASIC, FPGA) three under consideration. source code experiments available at https://github.com/SmartEngines/bipolar-morphological-resnet.