Periodic Reporting for period 2 - MacSeNet (Machine Sensing Training Network)
Reporting period: 2017-01-01 to 2018-12-31
We encouraged a reproducible research approach, through open publication of papers, data and software, and fostered an entrepreneurial and innovation-oriented attitude through exposure to SMEs.
In the research we undertook, we went beyond the current sparse representation and compressed sensing approaches, to develop new signal models and sensing paradigms.
We developed new robust and efficient Machine Sensing theory and algorithms. We applied these methods to real-world problems, through work with non-Academic partners, and disseminated the results of this research to a wide range of audiences, including through publications, data, software and public engagement events.
Our Fellows have produced results across 5 research areas:
In Core Theories and Algorithms they have developed a number of new methods, including using approximations where trying to find exact solutions would be intractable; new efficient methods that are scalable to large problems; a special type of averaging for fast optimization algorithms that is better than traditional linear estimators; and speeding up algorithms by “compressing” large-scale problems into smaller problems, and exploiting the structure of the datasets.
In Advanced Brain Imaging and Analysis they have produced algorithms to: improve clinical diagnosis when using a reduced MRI scan time for patient comfort; improve the speed and results of the post-processing of MRI scans helping clinicians to better understand the scan results; improved the models when working with fMRI which measures the flow of blood in the brain so that scientists and doctors can see which areas of the brain are working; improved the way researchers can look at EEG and fMRI results in combination to understand how the brain works.
In Inverse Imaging Problems they have developed techniques to remove image noise from MRI scans, deblur images including noisy natural images and produce better noise models for all types of image denoising.
In Audio Machine Sensing they have studied how to remove non-linear noise such as clipping and quantisation, detect audio events from data with little training and separate the singer from the music on an old jazz track.
Going beyond traditional signals they have found a new, graph-based method of learning space-time signals which has been applied to natural language modelling and brain network analysis and a method for detecting anomalies in large graphs allowing for event detection using the dynamics of Web and social networks.
These results were published in 9 journal articles, 61 conference papers and 6 technical reports. They worked directly with industry to share their techniques with those who will use them, leading to a patent application on image denoising techniques, and they have presented their research to the wider public via outreach activities.
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; A special type of averaging for SGD with conditional exponential families potentially leads to convergence not to a suboptimal, but to the optimal solution; sketched gradient algorithms and improved variants of state-of-the-art accelerated stochastic variance-reduced gradient and randomized coordinated descent methods.
Optimizing the acquisition order of MRI samples to preserve image contrast for accelerated scanning; Accelerated Cartesian Magnetic Resonance Fingerprinting (MRF) using echo planar imaging that can be used to perform high resolution MRF; Accelerating processing and analysis of MR Spectroscopic Imaging data by introducing machine-learning and sparsity based methods that perform better than conventional solutions; a spatio-temporal denoising approach based on total variation regularization for arterial spin labelling MRI; a deep convolutional neural network based framework for the direct estimation of pharmacokinetic parameters from under-sampled DCE-MRI stroke acquisitions; Information Assisted Dictionary Learning to incorporate external information for enhanced robustness against miss-modelling that introduces a new sparsity constraint which does not require parameter tuning; A higher order unfolding which uses the Block Term Decomposition for more noise-robust fMRI Blind Source Separation; Fusion of EEG and fMRI via soft coupled tensor decompositions.
An algorithm for phase restoration which provides state-of-the-art results in the restoration of InSAR and MRI interferograms; a multi-resolution windowed Fourier filter for phase denoising that outperforms the state-of-the-art InPhase denoising algorithms; a framework tailored to class-specific image deblurring which outperforms several generic techniques; a framework which handles deblurring for natural images with high noise levels; improved iterative sparse recovery approaches; a framework able to practically combine a CNN with a non-local self- similarity based filter, achieving state of the art performance in image denoising.
Sparse coding and dictionary learning algorithms for nonlinear measurements such as audio clipping and quantization were developed; a new method for audio event detection from weakly labelled, minimal data; a deep learning architecture for robust singing voice separation and an algorithm that allows us to examine how deep learning modules process audio data for audio separation.
A deep learning model able to learn spatio-temporal signals on graphs which shows state of the art results on natural language modelling and brain network analysis; A fast and scalable method for anomaly detection in the dynamics of large time-evolving graphs with applications to Web and social networks; Unsupervised pre-training for Echo State Networks that outperform generic approaches relying on random initialization; Unsupervised event detection using the dynamics of Web and social networks.
These advances will lead to: faster MRI scans so greater patient comfort; a better understanding of the human brain; clearer photos and images for medical and industry applications; better audio quality; machines that can listen and help humans who can’t or aren’t there to hear. In addition, the advances in core algorithms and beyond traditional signals underpin these technologies, giving new ways to look at highly interconnected data, like social networks.