Yanzhi Wang Receives Army Young Investigator Award to Bring Deep Neural Network Machine Learning to Mobile Devices

prof next to network rack

A Deep Neural Network (DNN) teaches a computer how to think like a human mind, both flexible and complex. The machine learning of DNNs has previously been thought to require computations and memory storage capacity too large for mobile delivery. To address this, Assistant Professor Yanzhi Wang, electrical and computer engineering, has been awarded a prestigious Young Investigator Program Award from the Army Research Office (ARO) on ultra-efficient, real-time DNN acceleration on mobile platforms. The ARO YIP is awarded to outstanding scientists beginning their independent careers to attract them to pursue fundamental research in areas relevant to the Army, to support their research in these areas, and to encourage their teaching and research careers.

“Previously the coding and compilations required of DNN performance and accuracy were too much for mobile,” says Wang.  “Our work enables the machine learning to automate the coding and reduce DNN storage by  up to 6,645x on the mobile platform, saving manpower and processing power, and sacrificing no speed or accuracy.”

Wang’s work achieves end-to-end mobile DNN connectivity previously only thought possible within the computational and storage capabilities of desktop devices. Using a unique methodology of model compression, compilation, and design, the research offers a flexible model for DNN machine learning on the mobile platform.

Wang and his team focused first on pruning and quantization of Neural Networks based on the ADMM (Alternating Direction Methods of Multipliers) framework. Pruning requires researchers to train dense algorithmic networks, trim out the less important connections, then retrain the compressed neural networks. Wang’s work combines a depth of pruning and quantization that makes DNN-level mobile storage viable.

The compiler, based on the ADMM solution framework, acts as a bridge from the data set compression to hardware application, allowing for acceleration of the DNN process. At theory, algorithm, compiler, and hardware levels, the research demonstrates the potential of accurate end-to-end data transfer in real-time.

View videos of sample results.

Wang’s research opens up unprecedented possibilities for mobile devices. “Extremely high resolution object detection and recognition will be achievable,” he notes. “I can envision real-time translations and question-answering, automated license plate detection, and immediate access augmented reality and virtual reality applications.”

With just a phone or tablet, individual soldiers in the field will be able to more accurately recognize friendly or non-friendly objects, day or night, and give drones and helicopters more fidelity in target mapping. “In some cases, the processing speed can be up to 50x faster with the same accuracy,” Wang notes. “For populations without reliable internet, if we can provide high performance processing power on hundreds of billions of devices we can make access more equitable and have a big impact.”

Wang came to Northeastern from Syracuse University in the fall of 2018 with the vision of bringing this level of connectivity to the world. Now he plans to expand on scientific application, perhaps applications to better protect and guide soldiers in the field.

“I’m happy and grateful to receive the award, the work for which was only possible with the help of the department chair and the strong support of Northeastern. This new funding will allow us to expand on what we’ve achieved, enabling DNN to be more widespread, and making new things possible for all of us.”

ARO is an element of the U.S. Army Combat Capabilities Development Command’s Army Research Laboratory.


Related News

Running Deep Neural Networks on Mobile Devices

Related Faculty: Yanzhi Wang

Related Departments:Electrical & Computer Engineering