Frederic Theunissen has pioneered many fundamental statistical and information theoretical approaches to neuroscience. Free access to many neuroscience software tools are available on our software page and our github.
Systems Analysis and Information Theory
|
We have always been interested in information theoretic approaches for studying the neural code (Borst and Theunissen, Nat Neuro, 1999). The use of information theory has the advantage that it is model independent: in other words, one can quantify the amount of information transmitted about certain stimulus parameters without having to extract the stimulus-response function of the neuron. Information theory can also be use to analyze the nature of the neural code; for example, to assess the importance of temporal patterns in spike trains or the potential synergy in a population code. One of the difficulties for the use of information theoretic measures is that it requires very large data sets, which often makes it unpractical in sensory physiology, in particular when the space of stimuli being explored is large. Our major contribution to the field has been to develop estimation methods to address this data limitation issue. In previous work, we had shown that single neurons could be modeled as inhomogenous gamma processes and that information values could then be estimated from these models (Hsu et al, J Neurosci, 2004). More recently, we have developed an approximation to deal with under-sampling issues of the stimulus space (Gastpar et al, IEEE Inf Theo, 2010).
|
|
Some of our early work in collaboration with J. Gallant (UC Berkeley) involved the development of a suite of algorithms that we also implemented for release to the neurosciences community at large. This work resulted in two software packages: STRFPAK and STRFLAB. The core routine of STRFPAK is based on a ridge regression algorithm that was customized for natural sounds and images (Theunissen et al, 2001). STRFLAB is a much more extensive set of routines that include additional regularization methods and solutions for the generalized linear model.
In parallel, we examined the effect of different types of stimulus pre-processing for the estimation of STRFs in audition. We found that pre-processing steps that included compressive non-linearities and automatic gain control increased the quality of the STRF fit but that the use of a wavelet decomposition was not superior to the use of a regular spectrographic decomposition (Gill et al, J Comp Neuro, 2006). We also tested a much more complex representation of the stimulus that was based on the probability of finding particular features given prior expectations and recent past. The use of this stimulus representation allowed us to describe the response properties of secondary auditory neurons that failed to be described which the more classical STRF approach (Gill et al, J Neurophys, 2008). Finally, we also developed methods that allowed us to compare the STRFs obtained from different stimulus ensembles. To do so, one find a common sub-space for the normalization and regularization step in the multiple linear regression. We used this method to compare the STRF obtained when the auditory system is processing song to those obtained when the auditory system is processing matched synthetic sounds that lacked the behavioral meaning (Woolley et al, J Neurosci, 2006).
In parallel, we examined the effect of different types of stimulus pre-processing for the estimation of STRFs in audition. We found that pre-processing steps that included compressive non-linearities and automatic gain control increased the quality of the STRF fit but that the use of a wavelet decomposition was not superior to the use of a regular spectrographic decomposition (Gill et al, J Comp Neuro, 2006). We also tested a much more complex representation of the stimulus that was based on the probability of finding particular features given prior expectations and recent past. The use of this stimulus representation allowed us to describe the response properties of secondary auditory neurons that failed to be described which the more classical STRF approach (Gill et al, J Neurophys, 2008). Finally, we also developed methods that allowed us to compare the STRFs obtained from different stimulus ensembles. To do so, one find a common sub-space for the normalization and regularization step in the multiple linear regression. We used this method to compare the STRF obtained when the auditory system is processing song to those obtained when the auditory system is processing matched synthetic sounds that lacked the behavioral meaning (Woolley et al, J Neurosci, 2006).