A NumPy implementation of the bayesian inference approach of Deep Neural Networks as Gaussian Processes.

We focus on infinitely wide neural network endowed with ReLU nonlinearity function, allowing for an analytic computation of the layer kernels.

• Python 3
• numpy

Clone the repository

git clone https://github.com/MB-29/NN-gaussian-process.git

and move to the root directory

cd NN-gaussian-process

from nngp import NNGP

# ... 

regression = NNGP(
    training_data,              # Data
    training_targets,
    test_data,
    L,                          # Neural network depth
    sigma_eps_2=sigma_eps**2,   # Observation noise variance
    sigma_w_2=sigma_w_2,        # Weight hyperparameter
    sigma_b_2=sigma_b_2         # Bias hyperparameter
    )

regression.train()
predictions, covariance = regression.predict()

• A classification script for MNIST is provided in the file classify_MNIST.py. It relies on the additional requirement python-mnist available on pip.
• A 1D regression script is provided in the file 1D_regression.py. We obtained the following results.