Deep Learning Approach

About

Challenge Advances in Deep Learning have enabled computers to comprehend data in a variety of complex tasks. Yet, despite these advances, Deep Learning methods are often computationally intensive making them unsuitable for applications in portable or embedded Internet of Things devices. One example of this would be inferring behavior based on sensor data collected by a portable device. These types of scenarios often place challenging constraints on available storage and computational resources. A new approach is needed to allow for a more flexible application of Deep Learning methods that is compliant with reduced resource utilization based on the characteristics of the underlying platform.   Solution A new framework developed at Rice University allows a performance-efficient realization of Deep Learning on resource-constrained devices. Resource utilization can be greatly influenced by the dimensionality of the input samples. As such, a resource-aware data transformation approach is used to adaptively map the input data to a set of lower-dimensional subspaces. This results in a significant reduction in the network with minimal impact on inference accuracy. The data transformation is a pre-processing approach that is compatible with large dynamic datasets and acts to highlight the most informative aspects of the data. As a result, significant reductions in power and energy consumption, training time, latency, and memory footprint are realized. Benefits and Features   More efficient application of Deep Learning to reduce runtime and power and memory use by factors of 44 compared to other methods User-friendly API for implementation on different platforms Ability to preserve privacy of deep learning in Internet of Things devices Market Potential / Applications This approach has been tested in a variety of applications including indoor localization, activity recognition, hyperspectral imaging, and speech recognition, with up to 44-fold improvements over existing methods. However, it is applicable to any type of device with resource constraints, such as mobile or portable devices including cell phones, Internet of Things devices, or sensors. It would also enable scalable prototyping of smart sensing and learning tasks among distributed devices.