2020 Hardware implementation method for convolutional neural networks

In recent years, the use of Deep Neural Networks (DNN) has acquired great relevance due to their great capacity to extract useful information from large amounts of data. Their hardware implementation makes it possible to increase the speed of operation by performing parallel computing compared to approaches based on the use of microprocessors that use sequential computing architectures of the Von-Neuman type (software solutions). Convolutional Neural Networks (CNNs) are a type of DNN with a Feed-Forward connection, especially useful for shape recognition in images.

Thanks to the use of different techniques of approximate computing (ex. stochastic computing) in the implementation of the neural network a considerable reduction of the hardware requirements is achieved when compared to traditional binary logic. While area and energy requirements are reduced, accuracy is maintained. The neural network is implemented using a fully digital solution.