Deep Neural Networks (DNN) have become a critical component of tactical applications, assisting the warfighter in interpreting and making decisions from vast and disparate sources of data. Whether image, signal or text data, remotely sensed or scraped from the web, cooperatively collected or intercepted, DNNs are the go-to tool for rapid processing of this information to extract relevant features and enable the automated execution of downstream applications. Deployment of DNNs in data centers, ground stations and other locations with extensive power infrastructure has become commonplace but at the edge, where the tactical user operates, is very difficult. Secure, reliable, high bandwidth communications are a constrained resource for tactical applications which limits the ability to routed data collected at the edge back to a centralized processing location. Data must therefore be processed in real-time at the point of ingest which has its own challenges as almost all DNNs are developed to run on power hungry GPUs at wattages exceeding the practical capacity of solar power sources typically available at the edge. So what then is the future of advanced AI for the tactical end user where power and communications are in limited supply? Neuromorphic processors may provide the answer. Blue Ridge Envisioneering, Inc. (BRE) proposes the development of a systematic and methodical approach to deploying Deep Neural Network (DNN) architectures on neuromorphic hardware and evaluating their performance relative to a traditional GPU-based deployment. BRE will develop and document a process for benchmarking a DNN s performance on a standard GPU, converting it to run on near-commercially available neuromorphic hardware, training and evaluating model accuracy for a range of available bit quantizations, characterizing the trade between power consumption and the various bit quantizations, and characterizing the trade between throughput/latency and the various bit quantizations. This process will be demonstrated on a Deep Convolutional Neural Network trained to classify objects in SAR imagery from the Air Force Research Laboratorys MSTAR open source dataset. The BrainChip Akida Event Domain Neural Processor development environment will be utilized for demonstration as it provides a simulated execution environment for running converted models under the discrete, low quantization constraints of neuromorphic hardware.
Benefit: The proposed effort delivers a crucial first step in deploying edge Artificial Intelligence hardware platforms at the edge for both tactical and commercial applications. Because this proposal addresses a general framework for the evaluation and deployment of applied neural networks it is unconstrained but any particular set of use cases. Any end user that requires neural networks running under low power restrictions can benefit from the ability to systematically evaluate the efficacy of a model in order to understand the power versus performance trade they are faced with. In the tactical realm applications involving unmanned or unattended systems such as covert sensors/communications devices, sonabuoys, small platform UAVs, etc. that have limited available power can leverage neuromorphic hardware to ensure the AI capabilities deployed on them can operate as long and as efficiently as possible. Commercial opportunities in the areas of coastal/harbor surveillance are available. We became aware of the needs for harbor surveillance on our BRIZO SBIR effort (N193-A01-0647, N6833520F0095), in which we explored AI/ML for AIS anomaly detection. In cases where traditional AI/ML cannot differentiate anomalous behavior, Neuromorphic computing may help. This could be applied (re-trained via transfer learning) to harbor surveillance radar data, passive acoustic, or other relevant sensory input. This could significantly reduce congestion and improve collision avoidance. We would envision a base product aimed at passive sonar acoustic monitoring using a straightforward implementation of the technology, as well as market-based decision points to drive whether we invest in re-training the model for other modalities. The coastal surveillance market is projected to grow from USD 29.08 Billion in 2016 to USD 34.34 Billion by 2021. In much the same way, commercial fishing can benefit from advanced fish detection algorithms based on neuromorphic computing. Commercial fishing worldwide is a huge industry over $240 Billion. Much of that expense is in boats and fuel to travel to fishing grounds. If technology can help fishing boats be more effective, there would be a huge value proposition to the industry. Sales projections for a non-existent product in a global industry would be impossible to predict with accuracy, but for an industry that large, a market segment for a neuromorphic fish detector could easily be in the tens of millions of dollars. We would bring our product to market by partnering with an existing fish finder company, such as Furuno. Their current top-of-the-line product retails for $7,500, and we estimate that with sufficient value proposition, an additional 10-33% margin could be had. With up to $2,500 of revenue per sale, we expect we might see $1,000 through a licensing arrangement per sale. Sales of only 1,000 units would be $1M in annual revenue (sustaining), and we believe there is significant upside to this number.
Keywords: Event Based Processing, Event Based Processing, Power Characterization, Artificial Intelligence, Neuromorphic Processor, Edge processing, deep neural networks, Machine Learning, Performance Benchmarking
