Learning from minimally labeled data with accelerated convolutional neural networks

The main objective of an Artificial Vision Algorithm is to design a mapping function that takes an image as an input and correctly classifies it into one of the user-determined categories. There are several important properties to be satisfied by the mapping function for visual understanding. First, the function should produce good representations of the visual world, which will be able to recognize images independently of pose, scale and illumination. Furthermore, the designed artificial vision system has to learn these representations by itself. Recent studies on Convolutional Neural Networks (ConvNets) produced promising advancements in visual understanding. These networks attain significant performance upgrades by relying on hierarchical structures inspired by biological vision systems. In my research, I work mainly in two areas: 1) how ConvNets can be programmed to learn the optimal mapping function using the minimum amount of labeled data, and 2) how these networks can be accelerated for practical purposes. In this work, algorithms that learn from unlabeled data are studied. A new framework that exploits unlabeled data is proposed. The proposed framework obtains state-of-the-art performance results in different tasks. Furthermore, this study presents an optimized streaming method for ConvNets’ hardware accelerator on an embedded platform. It is tested on object classification and detection applications using ConvNets. Experimental results indicate high computational efficiency, and significant performance upgrades over all other existing platforms