Project description
Running deep neural networks on low-power IoT devices
The internet of things (IoT) and the rapid rise of artificial intelligence and machine learning has created a whole new set of challenges. One of these is the difficulty of running deep learning algorithms between diverse hardware platforms. This is an issue that has been largely addressed with workflows based on CPUs and GPUs. However, this is not the case with low-power devices like smartphones, cars, or watches on which deep learning inference has been gaining traction. The EU-funded hls4ml project will develop an open-software library that will automatically adapt deep neural networks to electronic circuits by utilising high-level synthesis tools and reducing resource utilisation.
Objective
With Deep Learning becoming ubiquitous in our life, running Deep Learning algorithms in real time on an heterogeneous set of hardware platforms is a pressing need in many aspects of our society. While traditional workflows based on standard CPUs and GPUs are established, Deep Learning inference on low-power devices (e.g. cars, smart phones, watches, etc) is gaining more attention. Typically, this would require strong background in electronic engineering to convert a neural network into a Digital Signal Processor. We propose to develop a complete open-software library to automatically convert Deep Neural Networks to electronic circuits, using High Level Synthesis tools. With a large basis of potential applications (e.g. autonomous cars, medical devices, portable monitoring devices, custom electronics as in the real-time data processing system of large-scale scientific experiments, etc.), the hls4ml library would assists users by automatising the logic circuit design as well as by reducing resource utilisation while preserving accuracy.
Fields of science
- engineering and technologymechanical engineeringvehicle engineeringautomotive engineeringautonomous vehicles
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineering
- natural sciencescomputer and information sciencesartificial intelligencemachine learningdeep learning
- natural sciencescomputer and information sciencesdata sciencedata processing
- natural sciencescomputer and information sciencesartificial intelligencecomputational intelligence
Keywords
Programme(s)
Funding Scheme
ERC-POC - Proof of Concept GrantHost institution
1211 GENEVE 23
Switzerland