ANC Workshop: David Sterratt and Sander Keemink Chair: Katharina Heil
What 


When 
Sep 22, 2015 from 11:00 AM to 12:00 PM 
Where  IF Room 4.31/4.33 
Add event to calendar 
vCal iCal 
Sander Keemink
Multivariable coding and decoding in the brain Neurons in the sensory parts of the brain are well known to be tuned for specific features, such as orientation in a specific visual region (receptive field), tone frequency, etc. Assuming a group of neurons codes for a single variable this coding is optimal and unbiased.
However, in reality neurons rarely code for only a single variable. Orientation tuned neurons, for example, are often also tuned for colour, spatial frequency, etc., and additionally modulated by orientations outside their receptive field and other contexts. It is still unclear how coding and decoding are affected by such multi variable tuning.
In this talk I show how decoding accuracy can already break down for some very simple models encoding just two variables, and how this heavily depends on the type of tuning used.
David Sterratt
Sampling from models of synaptic plasticity
Suppose we have a parameterised dynamic model of an experimental system in which some of the variables have been measured at certain time points. The loglikelihood of any set of parameters can be derived from the sum of the squared differences between the model variables and the experimental measurements at the time points in question.
Analysis of a range of systems biology models shows that around a point in parameter space, some directions are "stiff", i.e. moving in them causes rapid reductions in the loglikelihood, whereas others are "sloppy": the loglikelihood changes slowly in sloppy directions (Gutenkunst & al. 2007). Analysis of These directions can lead to insights into the relationship between parameters in the system. Sampling from parameter spaces produces families of time courses, and thus gives error bars on predictions made by the model.
I will investigate what insights "sloppy" analysis can give when applied to dynamic models of synaptic plasticity (e.g. Smolen & al. 2012). Due to the anisotropic geometry of the loglikelihood function, sampling methods that make use of its geometry will be more efficient than standard ones, and I will use the Riemann manifold Metropolis adjusted Langevin algorithm (MMALA, Girolami & Calderhead 2011) to sample from a model of synaptic plasticity.
Girolami & Calderhead (2011) "Riemann manifold Langevin and Hamiltonian Monte Carlo methods" J. Royal Statistical Soc. B 73, 123214
Gutenkunst & al. (2007) "Universally Sloppy Parameter Sensitivities in Systems Biology Models" PLOS Comp. Biol. 3, e189+
Smolen & al. (2012) "Molecular constraints on synaptic tagging and maintenance of longterm potentiation: a predictive model." PLOS Comp. Biol. 8