MicroLearn (2016 - 2019)
The Micropower deep learning project (MicroLearn) aims to push beyond the current power walls for deep learning and move toward micro-power deep learning. This requires working on algorithms, architecture, circuits as well as design methods for deep learning under extreme constraints: tiny energy buffers (batteries), miniature energy harvesters, low-power and low-cost logic and memory devices. Specifically, the project aims to study and develop the basis for a new 's generation of deep learning engines operating within a power envelope of a few mW. The project’s technical objectives are:
- Reducing the digital computational load, by curtailing the raw digital bandwidth produced by sensors through mixed-signal techniques which mix feature extraction with traditional analog-to-digital conversion.
- Increasing the energy efficiency of digital computational cores with massively parallel near-threshold accelerators, which retain flexibility and high performance on the application domain, while at the same time significantly reducing power with respect to general-purpose parallel instruction-set processors.
- Optimizing deep-learning algorithms to target ultra-low power hardware with limited computational and storage capabilities. Focus will be on robust training of classifiers with reduced numerical precision and weight compression techniques. Semi-supervised and unsupervised techniques for ultra-low power devices will also be explored in this context.
- Demonstrating algorithmic, circuit, and architecture innovations by developing real-life demonstrators of full-system prototypes based on advanced fabrication technologies.