# Aboleth¶

A bare-bones TensorFlow framework for Bayesian deep learning and Gaussian process approximation [1] with stochastic gradient variational Bayes inference [2].

## Features¶

Some of the features of Aboleth:

• Bayesian fully-connected, embedding and convolutional layers using SGVB [2] for inference.
• Random Fourier and arc-cosine features for approximate Gaussian processes. Optional variational optimisation of these feature weights as per [1].
• Imputation layers with parameters that are learned as part of a model.
• Noise Contrastive Priors [3] for better out-of-domain uncertainty estimation.
• Very flexible construction of networks, e.g. multiple inputs, ResNets etc.
• Compatible and interoperable with other neural net frameworks such as Keras (see the demos for more information).

## Why?¶

The purpose of Aboleth is to provide a set of high performance and light weight components for building Bayesian neural nets and approximate (deep) Gaussian process computational graphs. We aim for minimal abstraction over pure TensorFlow, so you can still assign parts of the computational graph to different hardware, use your own data feeds/queues, and manage your own sessions etc.

Here is an example of building a simple Bayesian neural net classifier with one hidden layer and Normal prior/posterior distributions on the network weights:

import tensorflow as tf
import aboleth as ab

# Define the network, ">>" implements function composition,
# the InputLayer gives a kwarg for this network, and
# allows us to specify the number of samples for stochastic
net = (
ab.InputLayer(name="X", n_samples=5) >>
ab.DenseVariational(output_dim=100) >>
ab.Activation(tf.nn.relu) >>
ab.DenseVariational(output_dim=1)
)

X_ = tf.placeholder(tf.float, shape=(None, D))
Y_ = tf.placeholder(tf.float, shape=(None, 1))

# Build the network, nn, and the parameter regularization, kl
nn, kl = net(X=X_)

# Define the likelihood model
likelihood = tf.distributions.Bernoulli(logits=nn).log_prob(Y_)

# Build the final loss function to use with TensorFlow train
loss = ab.elbo(likelihood, kl, N)

# Now your TensorFlow training code here!
...


At the moment the focus of Aboleth is on supervised tasks, however this is subject to change in subsequent releases if there is interest in this capability.

## Installation¶

NOTE: Aboleth is a Python 3 library only. Some of the functionality within it depends on features only found in python 3. Sorry.

To get up and running quickly you can use pip and get the Aboleth package from PyPI:

$pip install aboleth  For the best performance on your architecture, we recommend installing TensorFlow from sources. Or, to install additional dependencies required by the demos: $ pip install aboleth[demos]


To install in develop mode with packages required for development we recommend you clone the repository from GitHub:

$git clone git@github.com:data61/aboleth.git  Then in the directory that you cloned into, issue the following: $ pip install -e .[dev]


## Getting Started¶

See the quick start guide to get started, and for more in depth guide, have a look at our tutorials. Also see the demos folder for more examples of creating and training algorithms with Aboleth.

The full project documentation can be found on readthedocs.

## References¶

 [1] (1, 2) Cutajar, K. Bonilla, E. Michiardi, P. Filippone, M. Random Feature Expansions for Deep Gaussian Processes. In ICML, 2017.
 [2] (1, 2) Kingma, D. P. and Welling, M. Auto-encoding variational Bayes. In ICLR, 2014.
 [3] Hafner, D., Tran, D., Irpan, A., Lillicrap, T. and Davidson, J., 2018. Reliable Uncertainty Estimates in Deep Neural Networks using Noise Contrastive Priors. arXiv preprint arXiv:1807.09289.