This project contains python code for a model of narrative integration that allows simulation of human fMRI results
The experiments use multiple resources that should be installed:
The word embeddings are from Wikipedia2vec that should be installed https://pypi.org/project/wikipedia2vec/ and the embeddings must be downloaded http://wikipedia2vec.s3.amazonaws.com/models/en/2018-04-20/enwiki_20180420_100d.pkl.bz2
The reservoir model uses easyesn https://pypi.org/project/easyesn/
We require a modification to the easyesn source code, so that the esn.precit returns the reservoir internal state trajectory
Just modify the last line of the predict function` and change it from return Y.T to return Y.T, X
This is clearly explained at kalekiu/easyesn#12
The HMM segmenation model from Baldassano et al. (2017) Neuron, is included in Brainiak https://brainiak.org/
There is some preprocessed fMRI data (which originally comes from Uri Hasson's laboratory, on Datalad) http://datasets.datalad.org/?dir=/labs/hasson/narratives - note this is a link to the original data - not directly required for the current experiments
Experiment 1: Simple-segmentation - This runs the segmenation on reservoir states from reading the NYT and Wikipedia narratives. everything is there.
The state files from the reservoir are in the repository.
Experiment 2: compares 10 reservoirs to 10 subjects on not the fall. Requires 10 reservoirs – generated by Generate_reservoirs.ipynb, instead of using wiki2vec, we take the embeddings, in the file /CH2020_50/inputDataTestingIntact.npy, so need to upload this too, and edit Generate_reservoirs.ipynb/
For the 10 subjects create a mycore cnrs folder: they are in this repository: https://mycore.core-cloud.net/index.php/s/CURtzbnV2nkMshy
Experiment 3 on construction and forgetting: not-the-fall.txt not-the-fall-shift-2.txt
Experiment 4
Reservoir-method reservoir-dynamics-impulse