QIF is an annual programme that focuses on recognising, rewarding, and mentoring the most innovative engineering PhD students across Europe, India, and the United States.
The Europe programme rewards excellent young researchers in the fields of artificial intelligence and cybersecurity with individual prizes of $40,000, dedicated mentors from the Qualcomm Technologies team as well as the opportunity to present their work in person to an audience of technical leaders at the company’s HQ in San Diego.
Jim Thompson, chief technology officer, Qualcomm Technologies, Inc. “We are proud to see the next generation of inventors bring their work to fruition. The machine learning and cybersecurity solutions they propose can have a significant positive impact on society. Through the QIF programme we aim to break down some of the barriers that researchers encounter by providing funding and fostering collaboration.”
Michael Hofmann, director of engineering at Qualcomm Technologies Netherlands B.V.: “Choosing a winner has been especially challenging this year. The cutting-edge problems that the researchers are working on tackle exciting applications that range from autonomous vehicles, smart drones to security of touchscreens and interactions in VR. We’re honoured to support excellence in European research for the 12th year in a row”.
This year, two more universities joined the participants list, namely Technical University Berlin (TU Berlin) and Technical University of Darmstadt. They have been selected as a result of their academic excellence in the fields of machine learning and cybersecurity. The other participating schools submitting proposals are in Belgium (KU Leuven), the Netherlands (Delft University, Vrije Universiteit Amsterdam), Switzerland (ETHZ and EPFL), the United Kingdom (Imperial College, Cambridge University and Oxford University), Germany (Max Planck Institute for Informatics, Saarbruecken and Technical University of Munich) and Ireland (Trinity College Dublin).
After careful review, the following four winners were selected for their outstanding proposals:
Europe QIF programme winners:
James Allingham (University of Cambridge), supervised by José Miguel Hernández-Lobato, has been selected for his proposal: “Diversity-encouraging Priors for Cheap but Well-calibrated Uncertainty in Deep Learning”.
Standard deep neural networks are used very successfully in many applications, but these networks are often overconfident and remain unable to quantify uncertainty robustly and reliably in their predictions. Existing solutions for uncertainty quantification either require strong approximations and significant computational resources, or there is room for improvement regarding the accuracy and the calibration of the results.
James proposes a method that imposes priors which encourage diversity among predictions made by different parts of the model. With the right balance between encouraging diversity and getting accurate predictions, this method will enable the provision of computationally cheap but well-calibrated uncertainty estimates. He also plans to release an open-source library for training and deploying the resulting models. The project can have a significant positive impact on safety-critical or health-related applications such as cancer diagnosis or automated driving, where computational resources are limited.
Tim Georg Johann Rudner (University of Oxford), supervised by Yarin Gal, has been selected for his proposal: “A Fully Probabilistic Theory of Autonomous Decision Making”.
Machine learning applications remain largely absent from many high-stakes domains where autonomous interaction with the real world is required, such as autonomous driving or high-precision surgery. This is because deep learning systems are still not reliably safe or robust in rare edge cases and unexpected or otherwise challenging situations. However, this is the solution going forward.
Tim’s proposal is about developing a fully probabilistic framework for reinforcement learning to provide reliable and mathematically rigorous uncertainty quantification. In contrast to previous approaches, he proposes to treat both the learning process as well as the model components, such an agent’s policy, probabilistically. The approach will combine advances in probabilistic inference and modeling with probabilistic reinforcement learning. This will enable autonomous vehicles and machines to “know what they know” as well as what they don’t know, and therefore to operate more safely and reliably.
David Romero (Vrije Universiteit Amsterdam), supervised by Jakub Tomczak and Mark Hoogendoorn, has been selected for his proposal: “Continuous Kernel Convolution For ML”.
A core question in machine learning research is how to model long-term dependencies in an efficient manner. Long-term dependencies are ubiquitously found in high-dimensional data such as audio and video signals. Convolutional neural networks exhibit an important limitation for modeling such long-term dependencies because of their limited memory horizon also referred to as receptive field. To overcome this issue, David proposes continuous convolutional kernels in CNNs. The parametrization of the continuous kernel is achieved via a small continuous output MLP that generates kernel weights given positions as input. David proposes to explore the applicability of convolutional kernel CNNs for applications in which long-term interactions play an important role: high-resolution (autoregressive) generative modeling, (large-scale) imagery, and video processing.
Siddharth Gupta (EPF Lausanne), supervised by Abhishek Bhattacharjee and Babak Falsafi, has been selected for his proposal: "Rebooting Virtual Memory with Midgard."
Popular online services have extensive user bases that are generating data at an unprecedented rate. This drastic increase in the dataset sizes has led to servers with TB-scale memory capacity. This proposal addresses the problem that large memories in data centers outgrow Translation Lookaside Buffer (TLB) capacities and need more page table levels in Memory Management Units (MMUs), leading to large latencies near the CPU cores. Sid proposes to introduce an extra stage of address translation, allowing coarse-grain translation combined with access control near the cores and fine-grain translation to support fragmentation near the memories. The worst latencies are moved away from the CPU cores and are mitigated by the caches, which operate in the intermediate address space.
For more information about QIF, please visit:
https://www.qualcomm.com/research/university-relations/innovation-fellowship