This is the code used to create the DeepSpike submission (Artur Speiser, Srinivas Turaga, Jakob H. Macke) for the spikefinder challenge: http://spikefinder.codeneuro.org/

We train recurrent neural networks trained on the available annotated data. For details check Community-based benchmarking improves spike inference from two-photon calcium imaging data

To see how such networks can be trained in an unsupervised way see Fast amortized inference of neural activity from calcium imaging data with variational autoencoders

This algorithm uses the Theano and Lasagne deep learning libraries and is organized as follows:

• The DataPrep notebook shows how we preprocessed the raw challenge data.
• The Example_run notebook shows how to train a network on the first challenge dataset and plots some results.
• The SpikefinderSubmission notebook recreates our challenge submission (outcomment the respective parameter setting to train on each of the 5 datasets)

Code for recurrent cells was taken from Lasagne/Lasagne#629