Various signal processing algorithms applied on Ca recordings from mouse V1 neurons. The stimuli under consideration fall in two categories:

1. Drifting sinusoidal gratings of constant spatial frequency and 16 different drift directions
2. Natural scene movies from the van Hateren dataset (5 movies, 11 experiments)

The responses were collected by two-photon microscopy from mice injected with GCaMP6f via a virus. The responses to the gratings were sampled at 20 Hz and the ones to the natural movies at 50 Hz. The grating stimuli were repeated 10 times for each direction (160 presentations in total) and the natural movies 5-10 times (variable across experiments). This leads to 19200 recorded frames for the grating stimuli (including the periods when a gray screen was shown between trials), and 6000-20000 frames for the natural movies (again including the inter-trial gray screen).

The original dataset was a set of .mat structs, which contained raw fluorescence values for the cells studied, the sliding-window-normalised fluorescence values for the same cells, and the spike rates inferred therefrom. The sliding window computation was done for a window of length 400 frames, using the simple formula dFF = (F - F0) / F0 where F0 is the baseline fluorescence computed over this window. This normalised fluorescence is then used to infer the spiking rate using the Vogelstein deconvolution algorithm, which is also already available in the original dataset.

To make working with this data easier, the structs were read into NumPy arrays of the given format: (S, N, L, R) -> (stimulus index, neuron index, sample index, trial index) where the array now holds only the inferred spike rates and nothing else.

• Orientation tuning : tuning.py
• Hierarchical agglomerative clustering : clustering.py
• Reliability computation : reliability.py
  • Effect of clustering : currently hacked into reliability.py (TODO: fix)

• Using a nearest-neighbour approach : classify_single_trial.py

• train_mlp_response_model.py : to fit a multilayer perceptron to V1 responses using Theano.
• compute_simulated_responses.py : running the MLP on unseen videos from the breakfast dataset.
• hmm/data_to_htk.py : converting the MLP output from .npy to .htk
• hmm/create_mlf.py : splitting the simulated responses into a train and a test set
• hmm/run_hmms.py : running an HMM-based classification model on the simulated responses, to verify the discrimination ability of the simulated V1 output.
• Comparison with random projections : compute_random_projections.py, replacing the computation in compute_simulated_responses.py. The rest of the pipeline propceeds the same way.

• Linear regression-based reconstruction (reverse correlation)
  • Forward and backward models : reconstruction.py
  • Effect of regularisation : TODO

0. Ask me for the data, which is not a part of this repository.
1. Use mat_to_npy.py for converting the data from original mat files to the NumPy format.
2. make_movies.py to generate the stimulus movies in the case of gratings and to read them from mat files in the case of natural movies. This also generates edge representations of the same movies, with the edges computed in time (using frame differencing) and space (using a difference-of-Gaussians filter).
3. downsample_videos.py to downsample the movies using scipy.signal.decimate.

This should be enough to get the data in a ready state. From now on, it depends on the experiment you want to perform. Most experiments need a single file; the ones which have a sequence to them are described in some detail in the 'Experiments' section.