Community Research and Development Information Service - CORDIS

H2020

MacSeNet Report Summary

Project ID: 642685
Funded under: H2020-EU.1.3.1.

Periodic Reporting for period 1 - MacSeNet (Machine Sensing Training Network)

Reporting period: 2015-01-01 to 2016-12-31

Summary of the context and overall objectives of the project

The aim of this Innovative Training Network is to train a new generation of creative, entrepreneurial and innovative Early Stage Researchers (ESRs) in the area of measurement and estimation of signals using knowledge or data about the underlying structure. With its combination of ideas from machine learning and sensing, we refer to this topic as “Machine Sensing”.
All ESRs are being trained in research skills needed to obtain an internationally-recognised PhD; to experience applying their research in a non-Academic sector; and to gain transferable skills such as entrepreneurship and communication skills.
We encourage an open “reproducible research” approach, through open publication of research papers, data and software, and foster an entrepreneurial and innovation-oriented attitude through exposure to SME and spin-out Partners.
In the research we undertake, we go beyond the current, and hugely popular, sparse representation and compressed sensing approaches, to develop new signal models and sensing paradigms. These include those based on new structures, non-linear models, and physical models, while at the same time finding computationally efficient methods.
We develop new robust and efficient Machine Sensing theory and algorithms; methods for a wide range of signals, including: advanced brain imaging; inverse imaging problems; audio and music signals; and non-traditional signals such as signals on graphs. We apply these methods to real-world problems, through work with non-Academic partners, and disseminate the results of this research to a wide range of audiences, including through publications, data, software and public engagement events.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

MacSeNet has recruited 15 ESRs in 9 universities and 1 SME across 7 countries. We have run researcher development training in transferable skills and Software Carpentry courses to introduce Fellows to the tools to help them do better research. The project teamed up with the SpaRTaN ITN to host a Spring School offering lectures and tutorials on scientific topics; it was free for non-network researchers to attend. Recent training included research data management, data sharing, open access publishing and science communication, culminating in a project based sandpit where teams worked to solve a real world problem via Agile project management.
Our Fellows have made good progress in their individual projects, their main results are listed below.
In Core Theories and Algorithms: A method that uses approximate inference based on stochastic gradient and perturb-and-MAP to learn parameters of log-supermodular models; Score function extensions to sliced inverse regression problems (SADE and SPHD) which combine sliced inverse regression (SIR) and score function-based estimation; Sketched gradient methods GPIS and Acc-GPIS for constrained least-squares regression, including convergence analysis; Structured sequence modelling with graph convolutional recurrent networks.
In Advanced Brain Imaging and Analysis: Poisson disk subsampling resulting in better image quality for iterative compressed sensing reconstruction on MRI data; Application of non-local means for denoising of MR spectroscopic imaging data, applied to both 2D and 3D MRSI data; A reconstruction method that can give up to 8-fold acceleration through the use of only a subset of k-space measurements; A new method called Assisted Dictionary Learning which allows the incorporation of specific time courses as external available information; A higher order unfolding which uses the Block Term Decomposition for more noise-robust fMRI Blind Source Separation.
In Inverse Imaging Problems: Exploiting non-local self-similarity and sparsity of phase images for denoising of Interferometric Phase Images; A new method based on Gaussian mixture models (GMM) to use class-adapted image priors for blind image deblurring; A single image super resolution technique using iterative back-projection with a specially designed collaborative filter; A correlated coloured noise model in contrast to the usual uncorrelated white noise.
In Audio Machine Sensing: A novel sparse signal reconstruction technique based on simple statistics learnt from a training signal; Methods for Acoustic Event Detection in real-life environments, and a method for bird detection using masked Non-negative Matrix Factorisation; Methods based on additive spectrogram models and generalised time-frequency masking to improve music source separation.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

MacSeNet has published 24 papers to date and our Fellows have presented their research at conferences and workshops internationally.
Our Fellows have shown progress beyond the state of the art and have been examining the potential impact of their research in the following areas.
In Core Theories and Algorithms: A variational bound based on perturb-and-MAP: for parameter learning via maximum likelihood previous methods typically lead to a degenerate solution while the new one does not; Enhanced sliced score function models (SADE and SPHD) improve estimation in the population case over non-sliced versions; Sketched gradient methods GPIS and Acc-GPIS achieve improved computational efficiency; The graph convolutional recurrent networks can be applied to real-world applications such as predicting epidemic disease spreading.
In Advanced Brain Imaging and Analysis: Poisson disk subsampling with iterative compressed sensing outperforms both random and uniform subsampling while reducing the scan duration, offering the potential for a considerable reduction in MR scan duration, resulting in increased patient comfort and better quality scans of moving areas (e.g. cardiac MRI); Tensor based methods for improving the result of fMRI Blind Source Separation in the presence of high noise; The Assisted Dictionary Learning approach exhibits an enhanced robustness against miss-modelling compared with other dictionary learning techniques for fMRI data analysis.
In Inverse Imaging Problems: The class-adapted blind image deblurring method outperforms several generic techniques for blind image deblurring and can to handle high noise levels in the case of images known to contain text; The proposed single image super resolution technique shows that domain specific knowledge can, in some circumstances, outperform the popular deep neural networks.
In Audio Machine Sensing: The new sparse signal reconstruction technique shows improvement compared to baseline algorithms; The new method for bird detection can be used for different type of audio data and works well when not much data is available; The research on music source separation using additive spectrogram models and generalised time-frequency masking has been applied to singing voice separation.
In the near future our Fellows will have another soft skills training event, in science communication and entrepreneurship in Edinburgh in 2017; coinciding with the International Science Festival to allow ample opportunities for public engagement. We are also running a Summer School in collaboration with the SPARS workshop in Lisbon in 2017 and will open this up to external researchers. Through these events we will encourage our Fellows to develop into “T shaped People” capable of adapting in a fast-moving research field. As the Fellows start out on their secondments, many of them to SMEs, we will be encouraging them to experience the research industry and consider the impact their work can have in the real world.

Related information

Follow us on: RSS Facebook Twitter YouTube Managed by the EU Publications Office Top