Presentation title

Deep-Learning Oriented Smart Sensing for the Next Generation of Embedded Applications

Authors

Manuele Rusci, Francesco Conti, Alessandro Capotondi and Luca Benini

Institution(s)

Alma Mater Studiorum - Università di Bologna, Italy

Presentation type

Technical presentation

Abstract

Future embedded systems will require bringing advanced computational intelligence directly on the edge so that entire classes of complex applications, such as autonomous nano-robots, advanced industrial safety systems, and novel brain-inspired human-machine interfaces are enabled. A particularly interesting class of algorithms for applications of this kind is that of Deep Learning (DL) and Deep Convolutional Neural Networks (CNNs), which already enables specialized cognition-inspired inference from collected data for a variety of different tasks such as computer vision, voice recognition and synthesis, big data analytics. DL algorithms are currently typically applied far from where data is collected – however, deploying advanced functionality based on DL directly at the sensor interface is extremely attractive, as it could potentially enable vastly more intelligent applications. Even limiting to vision-based tasks, this could deliver implantable biomedical devices; completely autonomous nano-vehicles for surveillance and search&rescue; cheap controllers that can be ``forgotten'' in environments such as buildings, roads, and agricultural fields; smart industrial safety measures; and many other innovative applications. To bring DL near the sensor, it is necessary to fit tens of billions of operations within a tiny power budget and for many applications also stringent real-time constraints. In this talk, we focus on vision-based applications and report two approaches we are following in a vertically integrated fashion towards this goal. First, we use data representations that allow for the maximum density in terms of energy per operation vs quality of results, such as event-based or strongly quantized representations; in particular, we deploy Quantized Neural Networks (QNNs) and Binary Neural Networks (BNNs) can be deployed with significant energy efficiency gains both in commercial microcontrollers (STM32 Cortex-M class) and custom-designed hardware accelerators. Second, we show how using hierarchical processing, i.e. performing part of the computation by integrating mixed-signal processing directly on the sensor focal plane, it is possible to increase energy efficiency further. In particulare, it is possible to extract “smart events” instead of raw data and feed BNNs directly with these data, exploiting the capabilities of DL algorithms to the fullest, leading the way to vertically integrated Deep-Learning oriented smart sensing next generation of embedded applications.

---

Additional material

• Presentation slides: [pdf]

For more details on this presentation please click the button below: