BrightOwl Loader Loading

Post-doc in digital design for efficient embedded machine learning processors Leuven Belgium,  

Posted on : 01 May 2017

Project Description

FunctieSpecifically, the project iscurrently looking for a PhD or post-doc to on resource-efficientdigital implementations of machine learning processors, focused around Bayesianmachine learning (and more general: Probablistic Graphical Models),deep learning, and reinforcement learning.Recently deepneural networks, such as convolutional neural networks (CNNs) or longshort-term memory (LSTM) networks have gained enormous popularity in the signalprocessing community. In the micro-elecronics research domain this has sproutedattention on customized processors for efficient embedded deep neural networkinference. Our team has published several of these state-of-the-art processorsover the past few years.For the highercognitive layers, where often sensor fusion takes place, a second machinelearning paradigm is attractive: Bayesian learning and Probablistic GraphicalModels. These techniques enable to more smoothly inject expert knowedge intothe system, and reason about the sensed information. White such white box classifiers are attractive from a knowledge point of view compared to the black box  deep neural networks, their execution is still very computationallyintensive on traditional processors. And so far, no customized processors havebeen build for these workloads.With this project,we want to enable the power of Probablistic Graphical Models to embeddeddevices. This through custom processor design and hardwae accelerator design forboth online learning and inference tasks. This research will hece require acombination of algorithmic innovations (dealing with reinforcement learning andProbablistic Graphical Models) and hardware innovations (processor design,low-power optimization and chip tape-out). We have already prooven in thefield of Deep Learning Processors, that such hardware/software co-optimizationallows to save orders of magnitude on energy efficiency. With this project, wewant to achieve similar gains for the next emerging deep learning technologybeyond CNNs, DNNs, and RNNs, and enable the power of Probablistic GraphicalModels on embedded devices. In this project, we closely collaborate withresearchers from KU Leuven s machine learning group DTAI, as well as with UCLA smachine learning group.Profiel


3000 Leuven Belgium

Find a Job Find Candidates

Most Recent Searches

Most Famous Searches

You might also like