ANC Workshop: George Papamakarios and David Sterratt, Chair: Aleksej Stolicyn
What 


When 
Nov 21, 2017 from 11:00 AM to 12:00 PM 
Where  IF 4.31/4.33 
Add event to calendar 
vCal iCal 
George Papamakarios
Title:
Masked Autoregressive Flow for Density Estimation
Abstract:
Autoregressive models are among the best performing neural density estimators.
We describe an approach for increasing the flexibility of an autoregressive
model, based on modelling the random numbers that the model uses internally
when generating data. By constructing a stack of autoregressive models, each
modelling the random numbers of the next model in the stack, we obtain a type
of normalizing flow suitable for density estimation, which we call Masked
Autoregressive Flow. This type of flow is closely related to Inverse
Autoregressive Flow and is a generalization of Real NVP. Masked Autoregressive
Flow achieves stateoftheart performance in a range of generalpurpose
density estimation tasks.
David Sterratt
I hope to give a doublebill arising from two MSc projects.
1. A problem when inferring parameters of dynamic systems with noisy starting conditions
(work with Judy Borowski)
To infer the dynamics of biochemical reactions involving divalent cations, such as calcium, solutions are prepared containing the reactants and photolabile chelators, such as DMnitrophen, which "cage" the cation. A flash of UV light cleaves the chelator, uncaging the calcium, thus allowing the reactions to proceed. The progress of the reactions is measured by a calciumsensitive fluorescent dye. The experiments are repeated under multiple conditions by varying the concentrations of the reactants and changing the length of the light flash to control the amount of calcium uncaged.
The resulting set of fluorescence timecourses and known parameters (concentrations and known binding coefficients) is used to infer the unknown parameters in a model of the studied reactant's binding to calcium. In order to do this the entire system of reactants, including the photolabile chelator and fluorescent dye, is modelled as as a set of ordinary differential equations. The maximum likelihood estimate of the unknown parameters is found by minimising the goodnessoffit over all conditions between the experimental fluorescence and the fluorescence derived from the model. The Hessian of the goodnessoffit with respect to the parameters indicates the distribution of unknown parameters.
One problem faced is that the starting conditions are noisy: the length of the light flash predicts the amount of calcium, but with considerable noise around the estimate. If more calcium is released than predicted, the whole timecourse will be shifted, meaning that on average the simulated traces are poor fits to the data.
2. A proportional electoral system with single member constituencies
(work with Aisling Mac Ardle)
Due to the first past the post (FPTP) electoral system, general election results in the United Kingdom are disproportionate to the level of national support held by political parties. In contrast, proportional electoral systems, such as single transferable vote (STV) and party list systems, provide proportional results but require more complex voting. With the goals of (i) producing election results that reflect all votes cast by the electorate, (ii) having single member constituencies and (iii) maintaining a simple voting process, as in FPTP, we have developed a novel electoral system, the Concentrated Vote (CV). CV moves votes between constituencies from candidates with low vote shares to candidates from the same party with a higher vote share. We apply CV to the results of UK general elections from 1979 to 2017, showing that CV improves proportionality over FPTP. Voter acceptance of CV might depend on the geographical unit to which it is applied.