Text-mining neurosynth text corpus using Deep Boltzmann Machines

Subdirectories/files are as follows:

• Code: contains all code.
• Data: will contain preprocessed data (note this is not included due to storage constraints)
• Data/Raw: folder to store all collected abstracts (kept empty to due to storage contraints)
• Weights: contains fitted RBM/DBM models (only final DBM model stored here atm)

In order to reproce results or train DBM on neurosynth abstract corpus run the following scripts: 1. collectPaperText.py: collect abstract text, cleans text and stores in incidence matrix 2. PretrainLayer.py: selects architecture of each layer by comparing log-likelihood on validation dataset. Pretrains RBMs for each layer 3. FineTuneDBM.py: applies DBM training algorithm to pretrained layers

Note that step 1 can be skipped (cleaned data is provided in data folder). The final DBM employed is also saved in the Weights folder ("DBM_50_50_Full.p")

The following (non-standard) libraries are required:

• biopython.org - used to collect abstracts from PUBMED API
• textmining (http://www.christianpeccei.com/textmining/) - used to build term incidence matrix
• nltk (http://www.nltk.org/) - used for stopwords Note that these library are only required for the text collection and cleaning (collectPaperText.py) and can be ignored if this step is skipped

Below we provide some examples from the trained DBM model:

import os
import cPickle as pickle
os.chdir("Code")
from softmaxRBM import *
from dropoutRBM import *
from DBM import *
os.chdir("..")


# load in pretrained weights:
dbm = pickle.load(open("Weights/DBM_50_50_Full.p", "rb"))

# one step reconstruction:
print dbm.oneStepRecon("fear")
## ['fear' 'amygdala' 'response' 'activity' 'functional' 'responses' 'cortex' 'brain' 'regions' 'activation']

# embed a document:
text = "At the forefront of neuroimaging is the understanding of the functional architecture of the human brain. In most applications functional networks are assumed to be stationary, resulting in a single network estimated for the entire time course. However recent results suggest that the connectivity between brain regions is highly non-stationary even at rest."
print dbm.embedDoc(doc=text.split(" "))

# similarly we can embed a single word, for example:
print dbm.embedDoc(doc="fear")