aleximmer/bnn-predictions

Improving predictions of Bayesian neural nets via local linearization, AISTATS 2021

Improving predictions of Bayesian neural nets via local linearization

Accompanying code for the paper

Immer, A.*, Korzepa, M., Bauer, M.*, Improving predictions of Bayesian neural nets via local linearization, AISTATS 2021.

Predictions with Laplace-GGN

We simply replace the BNN predictive with a GLM predictive for the Laplace-GGN posterior approximation. The code provides means to compute the Laplace-GGN posterior (diagonal, KFAC, full) and use it to make predictions. In the following example, we construct the Laplace approximation from a trained model model. The 'kron' cov-type corresponds to KFAC. The inferred posterior enables sampling from the posterior predictive for a batch of input data X using posterior.predictive_samples_glm(X, n_samples=1000).

from preds.likelihoods import CategoricalLh
from preds.laplace import Laplace

# infer posterior with Laplace-GGN
lh = CategoricalLh()  # likelihood 
prior_precision = 1.  # prior
posterior = Laplace(model, prior_precision, lh)
posterior.infer(train_loader, cov_type='kron', dampen_kron=False)  # or 'full', 'diag'

# GLM predictions
glm_samples = posterior.predictive_samples_glm(X, n_samples=1000)
# BNN predictions
bnn_samples = posterior.predictive_samples_bnn(X, n_samples=1000)

For a running and worked example, see the two examples on regression and classification:

• classification example notebook
• regression example notebook

The two examples train a neural network until convergence and construct variants of the Laplace-GGN posterior approximation. The script plots the posterior predictive of the proposed GLM in comparison to the heavily underfitting BNN predictive. The underfitting can only be resolved by artificially reducing the posterior variance; using a different prior does not help as it fails across the entire range of values.

Regression example

The resulting plot compares the proposed GLM to the BNN predictive:

Classification example

In the following example, the BNN predictive underfits severely so that the contours are almost invisible:

Setup

We use python >=3.7. To install the dependencies, run pip install -r requirements.txt. To use a GPU, additional installations might be necessary (CUDA, etc.). Then, install the cml package with pip install .. Create the result directories mkdir run_results and mkdir runs.

# install requirements
pip install -r requirements.txt

# install the `preds` package
pip install .

# run tests (optional)
pip install pytest
pytest tests

bnn-predictions's Issues

Uncertainty quantification for glm predictive

Hi,

I was wondering if it is possible to include code to reproduce the aleatoric and epistemic uncertainty shown in figure B.3?

Thank you!

Error when compiling classification.py

Hi, I am trying to reproduce the paper results, however I got this error:

Traceback (most recent call last):
File "/Users/vince/BNN-predictions-main/experiments/classification.py", line 236, in
train=True, double=double)
File "/Users/vince/BNN-predictions-main/preds/datasets.py", line 167, in init
'waveform', 'satellite', 'digits', 'banana'], error_msg
AssertionError: invalid UCI classification dataset

If you have some thought how to fix it I will appreciate.
Thank you again!