Researchers to infuse DSP with neural network kernels to enhance performance of radar and communications

Jan. 6, 2020
SPiNN program will develop a new set of advanced neural network computing kernels that embed established physics-based mathematical DSP models.

ARLINGTON, Va. – U.S. military researchers are asking industry to develop a new set of advanced computational neural network computing kernels that embed established physics-based mathematical digital signal processing (DSP) models in military applications.

Officials of the U.S. Defense Advanced Research Projects Agency (DARPA) in Arlington, Va., issued a presolicitation last week (DARPA-PA-19-03-04) for the Signal Processing in Neural Networks (SPiNN) program.

The SPiNN project seeks to apply neural network signal processing to military communications and radar and make substantial improvements to these applications over what conventional DSP technology could offer.

Accurately communicating multi-dimensional complex modulated signals through non-ideal dynamic communication channels is critical to many U.S. military radar and radio communications applications, DARPA researchers point out.

Related: Expanding defense capabilities by applying deep learning techniques

Conventional DSP techniques recover distorted signals by executing dedicated processing physics models to mitigate impairments sequentially. They assume stationary channel models with Gaussian noise, and therefore have limited capability to process temporal dispersion, non-linear distortions, or interference and jamming artifacts.

These error-prone cascaded operations are incapable of discovering and mitigating unknown impairments beyond established simple channel models. Such approaches also are computationally intensive, with long latency and poor size, weight, power, and cost (SWaP-C).

Emerging machine-learning techniques promise a new generation of computational approaches with reduced compute complexity and latency. For example, recent advances in Deep Neuromorphic Network (DNN) demonstrate fast feed-forward inference for good accuracy once it is trained with high-quality data sets. Currently, DNNs are trained by data sets and do not use physics-based mathematical models.

Yet missing corner cases and other unseen events beyond the collected data sets often leads to insufficient or misinterpreted representations to cause critical mission failures.

Related: Abaco Announces Obox Evaluation Platform to Minimize Time, Cost and Risk of Developing Autonomous Military Platforms

To establish a reliable and accurate DNN model, remote cloud computing facilities are necessary to support a vast computational workload on a large volume of training data. This practice makes DNNs impractical for many U.S. Department of Defense (DOD) machine-learning models.

The Signal Processing in Neural Networks (SPiNN) program will develop a new set of advanced neural network computing kernels that embed established physics-based mathematical DSP models.

The SPiNN program will capitalize on established physics-based signal processing algorithms and mathematical tool kits to establish a set of trained, verifiable, accurate ,and efficient neural network kernels.

SPiNN seeks to transpose important linear and non-linear DSP function blocks such as Fast-Fourier Transform (FFT/iFFT), Multi-Input Multi-Output (MIMO), Matched Filter (MF), Kalman Filter (KF), trellis/Viterbi decoders, and error-correction codes with verifiable outcome and accuracy into pretrained and low latency neural network kernel representations.

Related: Sonar designers making the transition to commercial technology

These pretrained neural network kernels will be fine-tuned to real-world data, and should outperform traditional DSP models, which lack the inherent capability to capture and process events that are difficult to model.

SPiNN kernels first will build on these verified DSP model sets to establish pre-trained neural network discriminators to process the incoming data with known accuracy, and then will then combine the trained neural network discriminator block with a generative neural network block and adaptive learning transform layer to form a generative-adversarial network kernel.

This kernel will capture corner cases and extract additional hidden structures beyond the known DSP models. The resulting SPiNN adaptive neural network kernels will provide accurate signal processing in real time when facing a dynamic real world environment.

The SPiNN project will depend on the open exchange of data sets and common interface of emulator suites among performers. Once selected for SPiNN, performers must share data set and common interface to the emulation suites.

Related: Electronic binoculars from Northrop Grumman team to detect threats through brain activity

The first phase of the SPiNN program will develop and demonstrate signal processing kernels based on physics models and adapt the models with a transformer layer.

Each proposal should propose either a communications or a radar applications for these DSP kernels. Communications applications could involve mobile phones, internet of things (IoT), point-to-point link, or cognitive radios.

Radar sensing applications, meanwhile, could involve surveillance radar, moving target indicator radar, synthetic aperture radar, or automotive radars.

Companies interested should submit proposals no later than 31 Jan. 2019 to the DARPA BAA Website at The project should begin on 1 April 2020.

Email questions or concerns to Young-Kai Chen, the DARPA SPiNN program manager, at [email protected]. More information is online at

Voice your opinion!

To join the conversation, and become an exclusive member of Military Aerospace, create an account today!