Event:

In this talk I will present our recent work on two important problems: modeling fixations and neural spike times. From a statistical point of view, both types of data can be described as point processes, and a principled way of measuring the usefulness of the corresponding models is to evaluate how well they predict unseen data. While for spike response models it is now common to optimize and compare them with respect to their likelihoods, differences between the likelihood of salience map models have not been studied to date.
I will first present a likelihood comparison of several popular saliency-based models of fixations (joint work with Matthias Kuemmerer). We show that differences in performance between these models are mostly due to differences in modeling the image-independent center-bias. After subtracting out this bias, there is about 1 bit/fixation of total information provided by the structure of the image about where people look under “free viewing" conditions. The best models explain up to 25% of this information and we dissect this information in terms of local spatial frequency content. In addition, we developed time-dependent saliency maps which extract a similar amount of 0.25 bits/fixation from the correlations between subsequent fixations.
In the second part of the talk I will present the spike-triggered mixture (STM) model which is a generalization of some frequently used generalized linear models (joint work with Lucas Theis). The STM model enables us to automatically extract complex stimulus-response relationships from recorded data. We show that our model can lead to substantial quantitative and qualitative improvements over generalized linear and quadratic models, which we illustrate on the example of primary afferents of the rat whisker system.