PI: Witten, Daniela (contact); Buice, Michael
Institution: University of Washington
Title: Models and Methods for Calcium Imaging Data with Application to the Allen Brain Observatory
Fast non-convex deconvolution of calcium imaging data, applied to the Allen Brain Observatory
PIs: Daniela Witten (University of Washington) and Michael Buice (Allen Institute for Brain Science)
The Allen Brain Observatory is an unprecedented survey of neural activity in the mouse visual cortex, recorded using high-throughput two-photon calcium imaging in the awake mouse. It consists of data from nearly 60,000 cells from six areas of the mouse visual cortex, 13 Cre lines, and 4 layers, recorded from over 200 mice in 432 sets of three experimental sessions, as the mice were exposed to artificial (static and drifting gratings, locally sparse noise) and natural (images and movies) stimuli. The data are publicly-available at http://observatory.brain-map.org/visualcoding/. We have developed a fast new algorithm for deconvolving this calcium imaging data --- that is, estimating the times at which each neuron spikes --- using an l0 penalization approach. Our publicly-available software is described at https://jewellsean.github.io/fast-spike-deconvolution/. The results of applying our deconvolution algorithm to all 60,000 cells in the Allen Brain Observatory can be easily accessed through the Allen Software Development Kit, described here: https://allensdk.readthedocs.io/en/latest/. Together, the Allen Brain Observatory combined with our new software for spike deconvolution constitutes a valuable publicly-available resource for the study of visual coding, as well as a generalizable tool that can be applied to other calcium imaging data sets.
Grant #: EB026908
BRAIN Initiative Alliance video - February 2019: A Statistician and A Neuroscientist Walk into a BRAIN Grant
Link to Data/Model Reuse abstract
Slide for 2021 BRAIN Awardee Welcome Meeting:
2021 Brain PI Meeting
Link to Poster: No associated poster.
CNN MouseNet: A biologically constrained convolutional neural network model for mouse visual cortex
Iris Shi, Bryan Tripp, Eric Shea-Brown, Stefan Mihalas, Michael A. Buice
Convolutional neural networks trained on object recognition derive inspiration from theneural architecture of the visual system in primates, and have been used as models ofthe feedforward computation performed in the primate ventral stream. In contrast tothe deep hierarchical organization of primates, the visual system of the mouse has ashallower arrangement. Since mice and primates are both capable of visually guidedbehavior, this raises questions about the role of architecture in neural computation. Inthis work, we introduce a novel framework for building a biologically constrainedconvolutional neural network model of the mouse visual cortex. The architecture andstructural parameters of the network are derived from experimental measurements,specifically the 100-micrometer resolution interareal connectome, the estimates ofnumbers of neurons in each area and cortical layer, and the statistics of connectionsbetween cortical layers. This network is constructed to support detailed task-optimizedmodels of mouse visual cortex, with neural populations that can be compared to specificcorresponding populations in the mouse brain. Using a well-studied image classificationtask as our working example, we demonstrate the computational capability of thismouse-sized network. Given its relatively small size, MouseNet achieves roughly 2/3rdsthe performance level on ImageNet as VGG16. In combination with the large scaleAllen Brain Observatory Visual Coding dataset, we use representational similarityanalysis to quantify the extent to which MouseNet recapitulates the neuralrepresentation in mouse visual cortex. Importantly, we provide evidence that optimizingfor task performance does not improve similarity to the corresponding biological systembeyond a certain point. We demonstrate that the distributions of some physiologicalquantities are closer to the observed distributions in the mouse brain after task training.We encourage the use of the MouseNet architecture by making the code freely available.