HD-RNN Applications

Increases in computational power, data throughput, and device miniaturization have contributed to remarkable proliferation of sensors, and the number and resolution of these sensors continues to increase at a much higher rate than the data can be utilized. Today, analysis of these data feeds is predominantly performed painstakingly by hand, severely limiting the usefulness of these hardware investments. Since increasing manpower to properly exploit this information is prohibitively expensive, partial automation is the only viable solution. Deeply layered machine learning is a good fit for the application of computer-aided inference techniques to real-world scenarios.

Machine learning focuses on algorithms and architectures to train computers to perform tasks that humans find repetitive and for which there is insufficient manpower to process the volume of incoming data. Machine learning already garnered success in applications such as recognizing patterns in DNA, facial recognition, license plate reading, voice recognition systems, and bank fraud detection.

Despite such advances, existing machine learning models and architectures did not scale well to massive data sets; they did not work in real time; and the variety of tasks a single system could be used for was extremely limited. Each problem domain and data modality had to be specifically trained for each task. Systems were very limited in the information they could provide and represent, and the resulting system did not easily integrate into other systems. In addition, training the system was a black art: it was very easy to over-train systems to detect nonexistent patterns in noise, and real-world implementations suffered from rampant false positives and false alarms.

HD-RNN, as a novel implementation of deeply-layered machine learning, addresses the shortcomings of shallow, stove-piped systems of the past and traditional machine learning systems. HD-RNN based solutions are able to adapt and respond in a biologically-inspired way; different levels of the learning hierarchy correspond to different abstraction capabilities which can be used for robust classification; the architecture does not depend heavily on the modality of the data; and systems are scalable to very large data sets. In addition, HD-RNN easily integrates with other systems in a variety of different configurations, ranging from small footprint (on-board) to heavy-duty and high-performance server installations.