Sulaiman Sadiq, from the Arm-ECS Research Centre, presented a seminar on “Enabling Imagenet-scale Deep Learning on MCUs for Accurate and Efficient Inference”, in association with the Center for Spatial Computational Learning.

The recording of the seminar is available on YouTube.

Abstract: The recent success of Deep Neural Networks (DNNs) and the proliferation of connected devices in the Internet of Things (IoT) has led to the development of the field of TinyML which aims to develop models and frameworks that can process data locally on resource constrained IoT devices. Often, these devices host microcontrollers (MCUs), with limited internal memory in the order of kilobytes, and a single core for processing. Conventional approaches in TinyML focus on achieving high accuracy by deploying the largest deep learning model with highest input resolutions that fit within the size constraints imposed by the MCUs fast internal storage and memory. In this talk, we show that models derived within these constraints suffer from low accuracy and, surprisingly, high latency. We present the TinyOps framework that opens a new efficient design space for MCUs using larger external memories with operation partitioning and DMA based overlaying. We demonstrate the strength of the TinyOps design space by comparing efficient models from this space with state-of-the-art models derived from the internal memory design space.

Biography: Sulaiman is a PhD student at the University of Southampton under a MINDS CDT studentship in collaboration with ARM Ltd. He holds a BE in Electrical Engineering and two MSc’s in Computer Engineering and Artificial Intelligence respectively. He also has 6 years of experience in industry working on real-time embedded systems in the R&D sector. His current research focuses on enabling deep learning inference on microcontroller devices by optimising across the software stack through efficient inference software and model design.