SpykesML

How to build encoding models with machine learning methods

This repository accompanies "Modern Machine Learning Far Outperforms GLMs at Predicting Spikes"[https://doi.org/10.1101/111450].

Here you can find a Python class MLencoding that you can use for quickly making encoding models. For a tutorial on how to use MLencoding, see the notebooks folder.

The notebooks folder also has a "standalone notebook" that demos how to use some ML methods without our fancy class.

Currently implemented methods:

• GLM
• 2-layer feedforward net
• Random forest
• xgboost
• LSTM With
• k-fold cross-validation
• spike and covariate history options and more!

Installation

In a terminal:

git clone https://github.com/KordingLab/spykesML
cd spykesML
python setup.py install

Quick how-to:

Build the encoder:

model = MLencoding(tunemodel = 'xgboost')
print(model.params)

Fit and predict to some data:

model.fit(X_train, y_train)
Y_hat = model.predict(X_test)

Perform k-fold cross-validation:

model.fit_cv(X,y)

Use spike and covariate history as inputs:

xgb_history = MLencoding(tunemodel = 'xgboost',
                         cov_history = True, spike_history=True,
                         window = 50, #this dataset has 50ms time bins
                         n_filters = 4,
                         max_time = 250 ) #in ms
xgb_history.fit_cv(X,y, verbose = 2, continuous_folds = True)

See the tutorial for how to define parameters and build new encoding models.

Dependencies

Basics

• numpy
• pandas
• scipy
• matplotlib

Methods

• sklearn
• pyglmnet (glm)
• xgboost
• theano (NN)
• keras (NN)