Stacked Autoencoders Using Low-Power Accelerated Architectures for Object Recognition in Autonomous Systems

This paper investigates low-energy consumption and low-power hardware models and processor architectures for performing the real-time recognition of objects in power-constrained autonomous systems and robots. Most recent developments show that convolutional deep neural networks are currently the state-of-the-art in terms of classification accuracy. In this article we propose to use of a different type of deep neural network—stacked autoencoders—and show that within a limited number of layers and nodes, for accommodating the use of low-power accelerators such as mobile GPUs and FPGAs, we are still able to achieve both classification levels not far from the state-of-the-art and a high number of processed frames per second. We present experiments using the color CIFAR-10 dataset. This enables the adaptation of the architecture to a live feed camera. Another novelty equally proposed for the first time in this work suggests that the training phase can also be performed in these low-power devices, instead of the usual approach that uses a desktop CPU or a GPU to perform this task and only runs the trained network later on the FPGA. This allows incorporating new functionalities as, for example, a robot performing runtime learning.