Dr. Anthony Quinn

Anthony Quinn is a senior academic (associate professor) in electronic engineering at Trinity College Dublin. He has had tenure there for twenty four years, and is a Fellow of the College since 2008. He has specialized for nearly thirty years in the development of theory and methods in statistical signal processing. He has an international reputation in Bayesian methodology, i.e. fully probabilistic methods for coping with uncertainty, inspiring optimal design and algorithm flows. He co-authored the first, and highly cited, book (Springer, 2006) on Bayesian methods for the design of signal processing algorithms that can be implemented efficiently (i.e. low energy and computational costs) on computers and specialized hardware. His algorithms have been applied in several practical contexts, notably in signal and image analysis for medical diagnosis and advisory system design. More recently, he has collaborated with his long-time partners in the Czech Academy of Sciences on important problems in Bayesian decision-making and design, with applications in distributed knowledge processing for sensor networks. This work has led to what is probably the first formal Bayesian definition of the transfer learning problem, a key concern in machine learning. He is developing new results in this area currently, as a one-year Fulbright visiting scholar at the renowned Michael Jordan group in Statistics at the University of California, Berkeley. He has recently secured an appointment as research scientist in the Department of Adaptive Systems at the Czech Academy of Sciences (2017-20). Professor Quinn is a committed educator in electrical engineering. He has originated and developed innovative courses in probability (which has been approved for the entire Junior Sophister engineering student cohort in his School) and in statistical signal processing (a module which he developed for the MAI Fifth Year programme in Electronic Engineering, and which was central to the success of the programme in professional accreditation). He has served in many senior roles professionally in his discipline and in his College. He named, lead-authored and developed the E3 and its Strategy in 2012, which received strong endorsement internationally. It has since been adopted as a priority infrastructure project by College. He serves recurrently, and for many years, on committees of several of the best international conferences and funding agencies. He was general chair of the Irish Signals and Systems Conference in 2011, and has acted as external examiner in Ireland and France.

  BAYESIAN INFERENCE   Bayesian signal and system identification   Bayesian signal processing in medicine and communications   Communication theory   Digital signal processing   Estimation   Image segmentation   Information theory   Probability theory   Signal Processing   Variational / stochastic approximations Bayesian inference

Research expertise: Bayesian statistical theory & methods; Bayesian nonparametrics for robustness & flexible design; optimal decision making & estimation; signal processing; transfer machine learning; distributed knowledge processing in sensor networks; information-theoretic criteria; Bayesian inferential approximation; design of recursive algorithms for real-time signal processing & digital receiver design; image analysis (texture modelling & segmentation); computer-aided diagnosis for medicine; advisory system design. Research summary: Prof. Quinn has specialized in Bayesian methods for signal processing, system analysis & machine learning for twenty nine years. He works at the interface between electrical engineering, probability & statistics, and is among the few internationally with grounding jointly in electrical/information systems analysis & design, and in advanced statistical methodology. His current position as one-year Fulbright visiting scholar in the renowned Michael Jordan group in the Statistics Department at UC Berkeley reflects this, as do his invited lectures (UC Berkeley, UC Santa Cruz and UW Madison to date) in 2017. Bayesian methods focus on fully probabilistic settings for information-bearing processes/systems. Prof. Quinn focusses on Bayesian answers to fundamental problems. Rather than applications as primary goal, he applies his methods selectively. Typical outputs are therefore (a) fuller understanding of the assumptions underlying classical & empirical solutions (what he calls the `auditing property' of Bayesian methods), and (b) new designs arising out of Bayesian analysis (what he calls the `prescriptive property'). The following is a selective list of the `fundamental problems' above: (i) Objective assessment of parametric model complexity, leading to optimal trade-offs between parsimony and prediction; many publications in model order inference, signal detection, threshold-free parameter estimation, and automatic rank determination (ARD) in matrix decomposition. (ii) Design of recursive computational algorithms for flexible classes of observation models. These Bayesian algorithms - which can be hosted by computing hardware - provide attractive trade-offs between cost and accuracy. This is the central focus of Professor Quinn's (widely cited) 2006 Springer monograph with a former student, and related papers. (iii) Nonparametric Bayesian solutions for flexible model design and model-robust inference. Here, the model itself is inferred, without finite degrees-of-freedom constraints. (iv) Randomized design for knowledge processing and transfer learning. A number of major recent papers with his Czech collaborators address machine learning and decision making problems, processing knowledge specifications broader than the crisp data sets of signal processing. Designs are randomized, thereby allowing exploration, and circumventing the need for optimization. These contexts have been explored in selective applications. Respectively: (i) ARD led to improved reconstruction of noisy functional scintigraphic images of the human kidney; (ii) a novel variational Bayes filter-bank provided excellent smoothing of impulsively corrupted speech; in telecommunications, an iterative symbol detector for a digital receiver yielded improved performance over state-of-the-art Viterbi algorithm; (iii) a nonparametric stopping rule for sampling from unknown distributions revealed a new non-uniform quantizer design; and (iv) an application in distributed knowledge processing for sensor networks yielded a new way of pooling distributions, equipping it with error bounds and providing improved prediction; a new Bayesian transfer learning mechanism allowed external data from a patient population to be exploited, improving design of radio-iodine therapy for individual thyroid cancer patients.