Achieving diversity in human vision is one of the major challenges for AI research. In the vast majority of cases, we are better than machines at understanding the world around us. But machines are catching up—slowly but surely.

Michael Felsberg and his co-workers test many of the solutions they develop in the vision laboratory on Campus Valla in Linköping. For instance, between the huge glass walls, autonomous drones and small self-driving cars equipped with advanced sensors and cameras are test-driven. But the actual brain in the computer vision is behind the lens.

“The camera is just a light sensor; it can’t do anything else. The actual work is done by the code and the software behind the camera. It’s the same with people: the eye registers the light and the brain does the work,” says Michael Felsberg.

Imitating human vision might seem easy at first glance. When AI research began, the feeling was that computer vision would be solved with a simple camera—maybe a project for the summer break. Now, almost 60 years later, general computer vision has developed into one of the most salient challenges in AI research.

We are in a fascinating era where even low-resource devices, such as Internet of Things (IoT) sensors, can use deep learning algorithms to tackle complex problems such as image classification or natural language processing (the branch of artificial intelligence that deals with giving computers the ability to understand spoken and written language in the same way as humans).

IMDEA Networks researchers Andrea Fresa (Ph.D. Student) and Jaya Prakash Champati (Research Assistant Professor) have conducted a study in which they have presented the algorithm AMR², which makes use of edge computing infrastructure (processing, analyzing, and storing data closer to where it is generated to enable faster, near real-time analysis and responses) to increase IoT sensor inference accuracy while observing latency constraints and have shown that the problem is solved. The paper “An Offloading Algorithm for Maximizing Inference Accuracy on Edge Device in an Edge Intelligence System” has been published this week at the MSWiM conference.

The main obstacle they have encountered in conducting this study is to demonstrate the theoretical performance of the AMR² algorithm and validate it using an experimental testbed consisting of a Raspberry Pi and a server connected through a LAN. “To demonstrate the performance limits of AMR², we employed fundamental ideas from linear programming and tools from operations research,” highlights Fresa.

Edge computing is an architecture in which data and processing are placed as close as possible to end users. In cloud computing, in contrast, the computations happen in centralized data storage and processing units in which algorithms enjoy the luxury of complex computations. Edge computing decentralizes such computations for reasons of privacy, latency, or energy efficiency. For his Ph.D. thesis, Emad Ibrahim used methods and practices from digital signal processing and machine learning to generate new solutions hosted on edge platforms.

Any edge device requires also a means of control. Acoustics from commercially off-the-shelf (COTS) sensor components can be used as a means for contactless control via in-air ultrasonic gestures and/or speech. The core brain for such control is a systematic combination of DSP and ML techniques.

One such application presented in Ibrahim’s work is a Virtual Proximity Detector (VPD) that can be used to robustly detect proximity in smartphones using solely the ultrasound band in the built-in speaker. Another application is detecting hand gestures using a small-form factor sensor array that is small enough to be embedded in consumer electronics such as smart speakers.