Arlington, Virginia – U.S. military researchers are calling on industry for scalable, robust and power-efficient analog neural networks to interface with the analog outputs of traditional sensors.
Officials at the Defense Advanced Research Projects Agency (DARPA) in Arlington, Virginia, on Wednesday released a broad agency announcement (HR001124S0022) for the Scalable Analog Neural Network (ScAN) project.
The ScAN program will develop a novel analog neural network that can directly interface with the analog output of traditional sensors and demonstrate three orders of magnitude power reduction over existing solutions.
A neural network is part of an artificial intelligence (AI) system that teaches computers to process data in the same way that the human brain does. It is a type of machine learning and deep learning that uses interconnected nodes to create adaptive systems that allow computers to learn from their mistakes and continuously improve.
Related: Artificial intelligence and machine learning in signal processing help intelligence analysts
The ScAN system demonstrates the inference capabilities of analog neural networks while eliminating the need for analog-to-digital converters at the raw sensor level.
Approaches such as in-memory computing use non-volatile memory in passive analog crossbar architectures that are inherently sensitive to process, voltage, and temperature variations, as well as parasitic and stochastic conductance variations. These approaches are susceptible to aging and drift, and require periodic tuning and retraining to maintain performance.
Instead, the ScAN program aims to use analog techniques to make neural networks more power efficient than today's digital techniques, while providing a high degree of inference accuracy, robustness to process, voltage and temperature variations, and scalability.
Today’s neural networks are typically digital, limited by power consumption for fast sensor processing, and can be too large, heavy, and power hungry for constrained sensor applications.
Related: Researchers plan to incorporate neural network kernels into DSPs to improve radar and communications performance
Recently, analog in-memory computing has been introduced into digital neural network systems in part to improve power efficiency, but it also increases the use of power-hungry A/D and D/A converters. These architectures are only reliable at small scales.
The ScAN program will demonstrate the inference capabilities of analog neural networks and develop novel analog neural network processing architectures and algorithms while eliminating the need for A/D converters at the raw sensor level.
ScAN has two technical areas: efficiently overcoming scaling limitations and short-term performance variations, and efficiently overcoming device-dependent variations and long-term performance changes.
Related: Incorporating Sensors and Signal Processing into Computing at Combat Speed
The ScAN program is a 4.5-year effort that also focuses on analog neural network design, simulation, and hardware development. Selected companies must use existing computing infrastructure or cloud-based services for all computing needs; they will not be permitted to acquire new computing hardware.
Interested companies should submit outlines by July 8, 2024, and complete proposals by August 8, 2024, on the DARPA Broad Agency Announcement Tool (BAAT) online at https://baa.darpa.mil/Public/SecurityAgreement.
If you have any questions or concerns, please contact DARPA by email. [email protected]For more information, visit https://sam.gov/opp/c6687528fcc54113a33256ec445cd0a2/view
