The research community gathered in this EDL program represents all the relevant academic disciplines, ranging from pattern recognition, embedded processor design, to high performance distributed computing. The applicants have demonstrated in the past to lead developments in their domain, and to be able to unite scientific excellence with industrial applications (spin-off companies, industrial support and impact). This excellence is further evidenced by an extensive track record of successful national (NWO, STW, RVO) and international (FP5/6/7, EU-Itea3, Artemis, Horizon2020, IARPA) projects.
Societal Impact
The DL revolution started around 2012 when DNNs could be realized by exploitng the compute power of GPUs with large public annotated data-sets (e.g. ImageNet). DL proved to be a real game changer, e.g., self-driving cars influence car manufacturers, but also chip developments. This time the expected impact will be much larger, with intelligent assistants everywhere, improving our capabilities and our quality of life in unprecedented ways. DL will penetrate almost any market; to name just a few examples:
• Transport/traffic management: collecting and analyzing detailed traffic situation.
• Media: (3D) video and image analysis, personalized content.
• Finance: Fraud detection, risk analysis and prediction of financial markets.
• Medical: image analysis, analyzing patient records, wearable devices, personalized medicine, etc.
• Manufacturing: visual quality inspection, release management, root cause analysis.
• Science: Discover life-changing answers faster: understand how cells respond to disease or genetic variations; understand the properties of innovative materials for clean energy and new semiconductors.
EDL results are generically applicable to these markets, as will be demonstrated by its use cases.
Economic Impact
The market for DL applications is huge, 1.77 Billion$ in 2022, growing 65% per year, with object/image recognition having its largest share. Since the Netherlands has already developed strong positions in Machine Learning, in ultra low power Embedded Systems, and in High Performance Computing, Dutch industry is at a pole position to grab a fair share of the huge market that is foreseen for DL applications and systems. EDL unites these Dutch communities, and connects research to industrial use cases. EDL results can be directly applied in products of our industrial partners, e.g in electronic microscopes by FEI, or road/environment monitoring equipment by TomTom, Vinotion and Cyclomedia. The table below gives the application perspective for the EDL industrial partners. Clearly an impressive list of applications is directly enabled by EDL developed technology and the demonstrated use cases
Roadblocks
The EDL program addresses seven deep learning roadblocks.
Deep Learning not efficient
- No labeled training data
- No real-time classification
- No Energy efficient platforms
Deep Learning not easy to use
- No understanding of what has been learned
- No expertise: SMEs don’t know how to use it
- No learning compute facilities for SMEs
- No data exchange betweeen universities & industry
Goals
The EDL program is structured into six research lines, R1 – R6, 3 for efficient learning and 3 for efficient computing.
| Learning | R1 : Data efficient DL | 100 x less training data |
| R2 : Efficient Multi-modal and Temporal DL | 25 x faster analysis | |
| R3 : DL verifiability, Accountability and Visualization | Understanding DL | |
| Computing | R4 : Efficient DL Processing Components | 5 x less energy & flexible processing HW |
| R5 : Efficient DL Architectures | 100 x more efficient DL architectures | |
| R6 : Distributed HPC for DL | Much lower DL deployment cost |