Séminaire LARIS - Lachlan J. Gunn
à 10h00 en salle du conseil à l'ISTIA
Le 12 janvier 2016
"Too good to be true : when Bayes turns abundant success into abject failure".
One measures and receives a positive result. One measures again, and again their suspicions are confirmed. Is it possible that this occurrence should in fact reduce our confidence? The dictum given in the title has existed for centuries, however in detection theory one normally chooses a threshold, below the test is negative, and above which it is positive.
Despite the intuitive simplicity of this approach, it neglects that there are often hidden states in which the system can be viewed as having failed. Including these low-probability states in the analysis can dramatically change the level of confidence provided by the combination of individual measurements.
We use Bayesian methods to analyse three problems of this type: the identification of the origin of an ancient Roman amphora, the search for criminals using identity parades, and the verification of primality
of large numbers for cryptograph, finding that high levels of confidence are more difficult to reach than is indicated by simple analyses.
Lachlan J. Gunn received his B. Eng. (Hons) and B. Ma. and Comp. Sc. (Pure) degrees from the University of Adelaide, Australia in 2012, receiving the 2012 J. Mazumdar Prize in Engineering and Mathematics, and four DSTO Scholarships in Radar Technology in the 2009–2012 period. In 2013 he was granted an Australian Postgraduate Award (APA), and is currently undertaking a Ph.D. under Derek Abbott and Andrew Allison. In 2014 he was awarded an Endeavour Research Fellowship by the Australian Government in order to undertake research into stochastic phenomena at the University of Angers, where is working within LARIS under François Chapeau-Blondeau, continuing a collaboration which has existed for many years between the University of Adelaide and the University of Angers.
His research interests include information-theoretic security and the use of stochastic signal processing for characterisation of nonlinear systems.