Our proposal Efficient Deep Learning (EDL) has been granted by the Dutch Science Foundation NWO. EDL brings together leading Dutch universities working on machine learning and efficient computation, and industry that is developing Big Data applications.

See NWO press release (in Dutch).

The Intelligent Vehicles group at TU Delft leads sub-project “Mobile Robotics” and will receive funding for two PhD positions related to automated driving. One PhD position involves research on deep learning for efficient mapping and localization, in collaboration with TomTom. The second PhD position addresses research on deep learning for 3D semantic traffic scene analysis, in collaboration with 2getthere.

Participating Universities

  • Eindhoven University of Technology
  • Delft University of Technology
  • University of Twente
  • University of Amsterdam
  • Vrije Universiteit Amsterdam
  • Wageningen University
  • Radboud University MC

Project Summary

Machine learning and in particular deep learning revolutionizes the fields of computer vision, language processing, medical technology, robotics, manufacturing, etc. Examples are text-to-speech conversion (Google Wavenet) and object recognition (ImageNet/COCO competition), both of which today can work at almost super-human accuracy. We have also seen recent successes in automated driving, with Tesla’s autopilot or NVidia’s self-driving car as notable examples.

Deep learning is a particular type of machine learning in which multiple layers of operations (e.g. convolutions) are stacked on top of each other. This forms a network which can learn a complex non-linear function for a specific task. Such networks have millions of parameters which can be automatically learned given large amounts of data and sufficient time to converge. In other words: deep learning can only succeed if enough data is available, and if training and inference are computationally efficient. New techniques such as fully convolutional networks, single-shot detectors, and residual and recurrent networks even increase the amount of parameters further. This calls for more efficient deep learning, encompassing both learning (offline or online) and inference phases.

Computational efficiency can be substantially improved by reconsidering all design levels, from application and algorithmic level, via network composition and structure, to mapping, processing architecture and hardware level. Typically at higher design (abstraction) levels higher gains can be achieved. Increased efficiency not only allows to deal with more complex networks, but also paves the way to use deep learning in embedded devices as used in various consumer applications, personal healthcare, domotica and automotive applications.

In this program we focus on various challenges, all addressing the performance and efficiency of deep learning; e.g. some involve particular machine learning techniques (e.g. continuous-time data, fine-tuning in the field, and spiking neural networks), others involve network structure and effective data and weight representations (e.g. network building blocks, learning network structure, compression, binary networks, sparse representations, and data fusion), yet others involve computationally efficient techniques (e.g. specialized deep learning hardware, distributed processing through crowd sourcing, exploiting data and processing locality, and accelerated GPU or FPGA implementations). The combined effect of solving these challenges enables to build vital and innovative applications in many domains like automated driving, public space mapping, healthcare, and electron microscopy.

In this Efficient Deep Learning program, we bring together four types of parties: 1) machine learning researchers who are advancing the state-of-the-art in deep learning, 2) companies which have access to huge data sets, 3) (embedded) system and HPC researchers who have the knowledge to make all of this computationally efficient to achieve advances in computational efficiency and 4) parties developing end applications that use deep learning. It is essential that such collaborations are formed in the Netherlands, since neither of these parties can perform all aspects of efficient deep learning by themselves, since the types or skills required are too diverse. With such a collaboration, the Dutch research community can truly advance the state-of-the-art in deep learning technology and make The Netherlands leading in developing innovative industrial and consumer applications exploiting deep learning.