Neural Data Analysis

We are interested in how neural circuits (primarily in sensory systems) perform the computations necessary to mediate behavior. To make progress on this question , we develop machine learning tools, statistical modeling approaches and data analytics solutions for making sense of the large scale high-dimensional data sets acquired by our experimental collaborators (visual cortex: Tolias lab; retina: Euler lab, barrel system: Schwarz lab, motor and olfactory system: Mathis lab). The following examples are representative of the type of work we do in this area:

System Identification wit Deep Learning Approaches

We developed scalable deep learning approaches to perform system identification in the mammalian visual system (Klindt et al. 2017, Ecker et al. 2019). Our work established that convolutional neural networks are state of the art for predicting neural responses to natural stimuli in primary visual cortex (Cadena et al. 2019).
system identification
Neural system identification with convolutional neural networks (CNN). First, the input is processed by CNN shared among all neuron. Next, each neuron's response is predicted by using a linear readout from the non-linear, shared feature space.

Retina Cell Clustering

We developed methods to cluster retinal cells into distinct cell types based on their responses to a battery of diverse visual stimuli, and found that the retinal output is organised into substantially more than 30 distinct functional channels (Baden et al. 2016). We used similar methods to classify bipolar cells and found that decorrelation of parallel visual pathways begins as early as the second synapse of the mouse visual system (Franke et al. 2017). Currently, we are working on extending this approach to investigate the functional organisation of the primary visual cortex (Ecker et al. 2019).
system identification
Neural system identification with convolutional neural networks (CNN). First, the input is processed by CNN shared among all neuron. Next, each neuron's response is predicted by using a linear readout from the non-linear, shared feature space.

State Space Modles

We used state space models for analysing the effect of anesthesia on the structure of population activity (Figure 1, Ecker et al. 2014). We showed that anesthesia-induced network state fluctuations lead to correlated variability and accounting for the network state recovers the awake noise correlation structure. This analysis allowed us thus to identify a major source of the controversy surrounding noise correlations in V1. In addition, we demonstrated that it is possible to infer internal brain signals l in real time from neural population activity, which has broad applications for example in studying the effects of attention or motivation on neural population activity.
Left: Noise correlations are higher in V1 of anesthetized animals. Middle: The difference is most pronounced for pairs with high average firing rate. Right: Schematic illustrating the Gaussian Process Factor Analysis model used for inferring the latent population state.

Inferring Firing Rates from Calcium Imaging Data

We developed a new algorithm for inferring firing rates from calcium imaging data as acquired in two-photon population imaging experiments based on flexible probabilistic models (Theis et al. 2015). Importantly, we show in the first extensive benchmark comparison to date on >50 cells for which ground truth is known that our algorithm outperforms all previoulsy published algorithms, even on datasets not seen during training.
Left: Example calcium trace and measured spike rate, as well as spike rates predicted by various algorithms. Right: Quantitative analysis of the performance of the different algorithms on four different datasets shows that our algorithm (STM) outperforms all previously published ones.

Key Papers

A. S. Ecker, F. H. Sinz, E. Froudarakis, P. G. Fahey, S. A. Cadena, E. Y. Walker, E. Cobos, J. Reimer, et al.
A rotation-equivariant convolutional neural network model of primary visual cortex
International Conference on Learning Representations (ICLR), 2019
#v1, #system identification, #microns, #convolutional neural network, #rotation equivariance
Code, URL, PDF, Data, BibTex

S. A. Cadena, G. H. Denfield, E. Y. Walker, L. A. Gatys, A. S. Tolias, M. Bethge, and A. S. Ecker
Deep convolutional models improve predictions of macaque V1 responses to natural images
PLoS Computational Biology, 2019
URL, DOI, BibTex

D. Klindt, A. S. Ecker, T. Euler, and M. Bethge
Neural system identification for large populations separating “what” and “where”
Advances in Neural Information Processing Systems 31, 2017
#convolutional neural networks, #system identification, #neural data analysis
Code, URL, PDF, BibTex

K. Franke, P. Berens, T. Schubert, M. Bethge, T. Euler, and T. Baden
Inhibition decorrelates visual feature representations in the inner retina.
Nature, 542, 439-444, 2017
URL, BibTex

T. Baden, P. Berens, K. Franke, M. Rezac, M. Bethge, and T. Euler
The functional diversity of retinal ganglion cells in the mouse
Nature, 529, 345-350, 2016
#retina, #clustering, #machine learning, #cell types, #ganglion cells
Code, URL, DOI, Dataset, BibTex

L. Theis, P. Berens, E. Froudarakis, J. Reimer, M. Roman-Roson, T. Baden, T. Euler, A. S. Tolias, et al.
Benchmarking spike rate inference in population calcium imaging
Neuron, 90(3), 471-482, 2016
#two-photon imaging, #spiking neurons
Code, URL, DOI, BibTex

E. Froudarakis, P. Berens, A. S. Ecker, R. J. Cotton, F. H. Sinz, D. Yatsenko, P. Saggau, M. Bethge, et al.
Population code in mouse V1 facilitates read-out of natural scenes through increased sparseness
Nature Neuroscience, 17, 851-857, 2014
#sparsity, #natural image statistics, #population coding, #v1, #two-photon imaging

A. S. Ecker, P. Berens, R. J. Cotton, M. Subramaniyan, G. H. Denfield, C. R. Cadwell, S. M. Smirnakis, M. Bethge, et al.
State dependence of noise correlations in macaque primary visual cortex
Neuron, 82(1), 235-248, 2014
#noise correlations, #gpfa, #population, #anesthesia, #macaque
Code, URL, DOI, PDF, BibTex

L. Theis, A. M. Chagas, D. Arnstein, C. Schwarz, and M. Bethge
Beyond GLMs: A Generative Mixture Modeling Approach to Neural System Identification
PLoS Computational Biology, 9(11), 2013
#generalized linear model, #spiking neurons, #mixture models
Code, URL, DOI, PDF, BibTex

T. Baden, P. Berens, M. Bethge, and T. Euler
Spikes in Mammalian Bipolar Cells Support Temporal Layering of the Inner Retina
Current Biology, 23(1), 48-52, 2013
#retina, #bipolar cells, #population coding

S. Gerwinn, J. Macke, and M. Bethge
Bayesian inference for generalized linear models for spiking neurons
Frontiers in Computational Neuroscience, 4, 2010
#bayesian inference, #generalized linear model, #spiking neurons
Code, URL, DOI, PDF, BibTex

J. H. Macke, S. Gerwinn, L. White, M. Kaschube, and M. Bethge
Gaussian process methods for estimating cortical maps
NeuroImage, 56(2), 570-581, 2010
#gaussian process
Code, URL, DOI, PDF, BibTex

University of Tuebingen BCCN CIN MPI