Live Demos

[Demo #1]

Live Demonstration: On-Device Learning for Domain Adaptation on Low-Power Extreme Edge Embedded Systems

Presenter Name: Cristian Cioflan

Presenter Affiliation: ETH Zurich, Germany


Keyword spotting accuracy degrades when embedded devices are exposed to noisy environments. On-site adaptation to previously unseen noise is crucial to recover lost accuracy, and on-device learning is required to ensure that the complete adaptation process can happen entirely on the edge device, thus respecting user privacy and security. In this work, we propose a complete on-device learning system targeting ultra-low-power microcontrollers, achieving real- time domain adaptation to unseen noises on always-on devices.

[Demo #2]

Real-time Image Classification Using An RNS-based DNN processor.

Presenter Name: Vasilis Sakellariou

Presenter Affiliation: Khalifa University, UAE


This demo showcases real-time image classification using a Residue Numbering System (RNS)-based CNN accelerator.  The prototype chip targets low-power edge-AI devices and considerably increases the energy efficiency (i.e. TOPS/W) of conventional systems. It features a number of innovative circuit and architecture-level optimizations which enable translating the efficiency of RNS regarding the implementation of the MAC operation into system-level performance gains. Silicon measurements show a 1.33× power reduction compared to the binary counterpart and considerably more energy efficient processing on CNN benchmarks compared to SOTA RNS-based CNN accelerators.

[Demo #3]

Secure All-Digital Neuromorphic Transceiver

Presenter Name: Mizan Abraha Gebremicheal

Presenter Affiliation: Khalifa University, UAE


IoT and Body Area Networks need protocols that meet strict power and space requirements. However, traditional serial transceivers struggle to meet those requirements. Though many alternative designs have been attempted, the presence of clock and data recovery (CDR) circuitry still poses challenges for reducing the size and power of the serial link transceiver. To address these challenges, we propose Spike-Coded Signaling, an asynchronous CDR-less communication protocol that translates data into pulse streams with information encoded in the count value of the pulses. Furthermore, a lightweight encryption protocol has been incorporated, leveraging the nature of SCS header and index packets to enhance security while minimizing the impact on data rates. In addition to ease of portability between technologies, the fully digital implementation results in a secure transceiver that is compact and energy efficient. In addition to the round-trip secure data transmission serial link, we will demonstrate a network of SCS transceivers using multiple Artix-7 FPGA boards.

[Demo #4]

Live Demonstration: Decoding Silent Speech: The Role of Magnetic Skins in Enhanced Speech Impairment Communication

Presenter Name: Montserrat Ramirez De Angel

Presenter Affiliation: KAUST, Saudi Arabia


Worldwide, approximately 70 million people suffer from speech impediments, meaning they cannot properly produce the sounds needed for effective communication. Current technologies often rely on sign language to facilitate communication; however, with over 300 distinct sign languages and only about 0.01% of the public fluent in these languages, communication barriers persist. To enhance natural communication for individuals with speech impediments, a novel approach focuses on mouth movements. The Assistive Magnetic Skin System (AM2S) tracks these movements using wearable sensors and deep learning architectures. This system employs biocompatible magnetic skin patches attached to the user's bottom lip. Through magnetic field sensors, it tracks changes in the magnetic field caused by mouth movements. These changes are then processed using a Convolutional Neural Network (CNN) integrated into an app, which visually displays the intended letters and words. This innovation enables individuals with speech impediments to engage in more natural conversations with others.