“Efficient Processing when Intelligence Moves to the Edge”

Continuous collection and processing of vast amounts of data is becoming more common with the advancement of wearable sensors, Internet of Things (IoT) devices and cyber physical systems. Despite their potential significant impact, intelligent mobile technologies face a number of challenges for use in daily life. These devices usually stream raw data to cloud computers for computation which leads to massive storage, significant transmission power consumption, real time constraints and privacy/security issues, thus processing at the edge of sensors is becoming increasingly preferred. Processing these sensor-level data requires a variety of signal processing and machine learning tasks that come at the cost of high computational complexity and memory storage which is overwhelming for these light weight and battery constrained platforms.

In this talk I will present research solutions that enable efficient processing of machine learning tasks to improve energy efficiency and throughout without sacrificing application accuracy for wide deployment of embedded/edge processing. First, I present DeepMatter a scalable framework across algorithms, architectures and hardware to design an embedded Deep Neural Net (DNN) accelerator. DeepMatter takes any number of raw sensor data for variety of applications and classifies the events within 92% accuracy and consumes very low power. Next I present two programmable domain-specific manycore accelerators namely PENC and BinMACthat achieve energy and speed efficiencies comparable to application-specific custom hardware through data-level and task level parallelization as well as customization of instruction-sets per-core and near-memory computing. I will show the efficiency of DeepMatter as well as PENC and BinMAC solutions for several application domains including multi-physiological processing for seizure and stress monitoring, tongue drive assistive device, air quality monitoring andvision-based situational awareness. The solutions derived at the intersection of algorithms, architectures and implementation allow designers to rapidly prototype and deploy the next generation of sophisticated and intelligent systems for efficient edge processing in extreme environments.