The code associated with the Bachelor's thesis "Dynamic coding in a large-scale, functional, spiking-neuron model". This code can be used to analyse (model) data. (See this repository for a Nengo model that can generate such data.) More specifically, it checks for either the regular or cross-temporal decodability of oriented gratings.

The meat of this code, which can be found in wolff.py and wolff_cross.py, is based on a paper by Wolff and colleagues (2017), called "Dynamic hidden states underlying working-memory-guided behavior".

Code for calculating decodability curves. The file has two main functions: similarity and similarity_p, where the latter is a parallellised and significantly faster version of the former. They have the following signatures:

def similarity(data, theta, angspace, bin_width):
    ...
    return (cos_amp, distances)

def similarity_p(data, theta, angspace, bin_width, num_cores):
    ...
    return (cos_amp, distances)

where the parameters indicate the following:

• data: a trials by channels by timesteps numpy array of EEG/model data;
• theta: a trials long numpy array of associated memory item angles;
• angspace: a numpy array of centres of the bins that will be used to bin data;
• binwidth: a value that indicates the bin width of the aforementioned bins;
• num_cores: the maximum number of cores similarity_p will use for its calculations.

The returned variables are:

• cos_amp: a trials by timesteps numpy array of decodability values, i.e. a decodability curve for every trial from data;
• distances: a trials by bins by timesteps numpy array of Mahalanobis distances. Mean-centring, convolving with a cosine and taking the mean across the bin dimension gives cos_amp.

Code for calculating cross-temporal decodability analyses (CTDAs), designed specifically to use GPUs. The main function to call is:

def cross_decode(data, theta, bin_width, sigma=None, device_i=None):
    ...
    return cross_cos_amp

where the parameters indicate the following:

• data: a trials by channels by timesteps numpy array of EEG/model data;
• theta: a trials long numpy array of associated memory item angles;
• binwidth: a value that indicates the bin width of the bins. The bin centres (angspace) are hardcoded;
• sigma (optional): a trials by timesteps by channels by channels numpy array of inverse covariance matrices. If not given, will be calculated by cross_decode;
• device_i: the number of which GPU to use. Not optional: cross_decode has been designed specifically for use with a GPU.

The returned variable is:

• cross_cos_amp: a trials by timesteps by timesteps array of decodability values, i.e. a CTDA for every trial from data.

Two Jupyter notebooks have been added that have been involved in personal use and can serve as examples. analyse_decode_mem.ipynb makes use of wolff.py to calculate decodability curves, while analyse_cross.ipynb uses wolff_cross.py to calculate CTDAs.

A Conda environment file, called dynamic-env.yml, has been added to quickly set up the required dependencies.