Overview

This repository contains the code for the analyses whose results are shown in Figure 9 panels A, B, D, F and G of Frechter et al. 2019.

Requirements

The main analyses are performed in Python 3.6.8. The required packages are listed in requirements.txt and can be installed using

pip install -r requirements.txt

The data required for the analyses is provided in the input_data.tar.gz compressed tar file, but can also be prepared from scratch using the instructions below. Doing so requires the gphys, physplitdata and physplit.analysis packages. The code has been developed and tested on macOS High Sierra and CentOS 7, and should run with little modification on similar *nix systems. Finally, gnuplot is used to make some diagnostic plots for the population decoding analyses.

Usage

Input Data

The data required for the analyses can be decompressed from input_data.tar.gz or built from scratch. Decompressing the file in the top level repository folder will yield an input_data folder containing the required files. Alternatively, it can be built from scratch by:

1. Creating a folder input_data at the top level of the repository.
2. Installing the R packages gphys, physplitdata and physplit.analysis packages.
3. Running prepare_data.R from the top level folder.
4. Running prepare_data.py from the top level folder.

AUC Analysis

The AUC analysis whose results are shown in Figures 9B and C can be performed by running compute_and_plot_auc.py. This will generate the box and whisker plot of Figure 9B, and a csv file of AUC values that can be used to plot Figure 9C. The plot and csv file are written to the output folder.

PCA Analysis

The PCA plots on the raw data shown in Figure 9A and for the divisively normalized data shown in Figure 9D are performed in compute_and_plot_pca.py. The plot is written to the output folder.

Identity and Category Decoding

Identity and category decoding is run as a parameter sweep over populations, number of classes, shuffles, etc. The sweep can be performed as follows:

1. A folder is created at the top level to contain the results of the runs. The name can be arbitrary; it's set to runs here.

2. The folder is initialized with a config.json describing the parameters of the runs to be generated.

3. The script gen_run_scripts.py is executed from inside the runs folder using

python ../gen_run_scripts.py

This will generate shell scripts of the form job*.sh, to be submitted to the SLURM job scheduler. Each script contains instructions for running a single instance of the analysis by calling the single.py script with the parameters of the run. In the absence of SLURM, the python call to single.py inside each job script can also be run manually. For example:

python -u ../single.py 0 PN 1 identity 0 1 0

See single.py for the meaining of the input paramters. Note that single.py uses gnuplot to generate a summary plot after the run is complete, so gnuplot should be available on the system.

4. Each job writes a number of files to the runs directory. These files are prefixed as e.g. seed12orig, indicating that the job used random seed 12 and used the original (unshuffled) data. The files are as follows:

   • seed(...).csv: Each row contains the accuracy at each time bin for a particular value of the linear SVM parameter C.
   • seed(...).pdf: The accuracy time course data in the csv file plotted using gnuplot.
   • seed(...)_summary.txt: A plain text file listing the parameters of the run as well as the peak accuracy achieved.
   • seed(...)_summary.p: The summary data as in ...summary.txt but stored as pickled python dictionary.
   • seed(...)_acc.npy: A num_C_values x num_time_bins x num_cross_validation_trials numpy array containing the decoding accuracy for each time bin, value of C, and cross validation trial.

5. The results stored in the runs folder are analyzed by process_results.py by running

python process_results.py runs

from the top level folder.

• This script will first assemble the pickled summary data generated by each run into a pandas dataframe. The dataframe will be written to runs/all_acc.df.p.bz2. Subsequent calls to process_results.py will look to load the dataframe from this file in favour of recomputing it, as the computation takes a few minutes.
• Once the data a processed, a number of performance plots are generated and written to the ./output folder. The most important is id_vs_cat_w_shuff.pdf which shows the identity and category decoding performance of each population as in Figures 9F and 9G in the paper. The remaining plots show the time course of accuracy for each population, as well as the best SVM parameter C the yielded the best performance at each time bin.

Questions and Comments

All questions and comments regarding the analysis should be directed to sina dot tootoonian at gmail dot com.