Optimizing deep learning for embedded systems
EmbeDL, a company based on the departments research on efficient deep learning has received funding to bring their innovation to the market. Potential customers are found in all areas using deep learning, from self-driving vehicles to the Internet of Things.
Chalmers participation in the Horizon2020 EU project LEGaTO (low energy toolset for hetergeneous computing) provided the basis for the innovation in EmbeDL. The aim for the EU project was to develop the next generation toolset for efficient heterogeneous computing, and EmbeDL is using the results to optimize deep learning models to make them faster and more energy efficient without compromising accuracy.
– Deep learning is a very powerful and successful technology which will drive the next generation of ubiquitous AI devices, for example in the Internet of Things, says Devdatt Dubhashi, professor in the Data Science and AI division at Computer Science and Engineering.
Devdatt Dubhashi is co-founder and chief scientific officer of EmbeDL, and sees the scaling down of large networks to allow them to run on small, inexpensive, heterogeneous hardware plattforms, as the major challenge. The core technology of EmbeDL involves a combination of algorithmic techniques to optimize both the structure and parameters of deep neural networks, so that they scale down dramatically and become much more efficient computationally, while ensuring sufficient accuracy.
Taking the product to market
Almi Invest recently announced that they will invest SEK 2.5 million in the company, and Chalmers Ventures, Stoaf III SciTech AB, Butterfly Ventures and Circus Future also participate in the share issue of SEK 7 million. The market for optimization of Deep Learning models is currently estimated at EUR 3.6 billion annually and is expected to grow explosively within the next few years.
– The investment will be used to take the product to market with initial focus on the automotive industry followed by IoT. In 2021, we will also launch the technology as a cloud-based platform, says Hans Salomonsson, CEO, who founded EmbeDL with Devdatt Dubhashi.
The company has managed to reduce the number of calculations in deep learning models by up to ten times. After optimization, the models can be used on inexpensive hardware and potentially result in large savings for companies deploying deep learning in embedded systems. The software can be used in all application areas where deep learning is used, from self-driving vehicles to the Internet of Things. A major challenge in deep learning and AI development is relatively heavy calculations, expensive hardware and difficulties in achieving real-time requirements.
– EmbeDL has clearly shown the value of its technology and Chalmers Ventures is now looking forward to assisting with the commercialization, says Jonas Bergman, Investment Director at Chalmers Ventures.
The industrial interest has been great from heavy technology companies, both nationally and internationally. The technology has been verified by industrial pilots in the automotive industry. Earlier this year, EmbeDL received an award from the prestigious international network HiPEAC for its commercial potential.
Press release from Almi Invest (Swedish)
Devdatt Dubhashi, professor, Data Science and AI division, chief scientific officer, EmbeDL.
Hans Salomonsson, CEO, EmbeDL
+46 730 63 28 37
The Department of Computer Science and Engineering is shared between Chalmers University of Technology and University of Gothenburg.