Qualcomm Technologies has just announced four winners of the Qualcomm Innovation Fellowship (QIF) for 2019. Among them are Mario Paulo Drumond and Kaicheng Yu, students at EPFL’s School of Computer and Communication Sciences (EDIC). They have been recognized by Qualcomm for their outstanding research proposals on emerging technologies.
Mario Paulo Drumond is mentored by Babak Falsafi, Professor at EDIC and founding director of EcoCloud. He figures among the QIF winners for his proposal “ColTraIn: colocated deep learning training and inference”. In this research, Mario proposes an accelerator design for co-located training and inference. The most significant part of the research is the use of Block Floating Point in tensor dot products, which account for most of the arithmetic in deep learning. There are two aspects of the proposed design. An exponent per tensor helps reduce memory to logic traffic and turn all tensor product arithmetic into fixed-point for higher logic density, while the employment of Floating Point for activations and remaining calculations improves accuracy.
Mario conducts his research at EDIC as part of the Parallel Systems Architecture (PARSA) group. Apart from his research experience, he has published several papers and worked on teaching assignments. He has also completed internships at Microsoft Research (Redmond and Cambridge).
Kaicheng Yu is supervised by Mathieu Salzmann and Pascal Fua, experts in computer vision and machine learning. He was selected by Qualcomm for his proposal “Robust Neural Architecture Search with Soft Weight Sharing.” Conventional network designs use a simple heuristic where different architectures are updated while sharing parameters. Conversely, Kaicheng introduces a new training scheme with soft weight sharing that enhances the use of neural architecture search (NAS) in deep learning. His proposal also presents a unified evaluation framework for the assessment of the algorithms’ ranking disorder.
QIF, now in its 8th year, is an annual program that focuses on recognizing, rewarding and mentoring the most innovative engineering PhD students across Europe, India, and the U.S. The awardees for 2019 were selected from diverse fields of research such as automated speech recognition, fingerprint recognition, 3D computer vision, and adversarial neural networks.