Skip to content

Bayesian Neural Networks

In this tutorial, we demonstrate how one can implement a Bayesian Neural Network using a combination of Turing and Flux, a suite of tools machine learning. We will use Flux to specify the neural network’s layers and Turing to implement the probabalistic inference, with the goal of implementing a classification algorithm.

We will begin with importing the relevant libraries.

# Import libraries.
using Turing, Flux, Plots, Random

# Hide sampling progress.

# Use reverse_diff due to the number of parameters in neural networks.
┌ Info: [Turing]: progress logging is disabled globally
└ @ Turing /home/cameron/.julia/packages/Turing/cReBm/src/Turing.jl:22


Our goal here is to use a Bayesian neural network to classify points in an artificial dataset. The code below generates data points arranged in a box-like pattern and displays a graph of the dataset we’ll be working with.

# Number of points to generate.
N = 80
M = round(Int, N / 4)

# Generate artificial data.
x1s = rand(M) * 4.5; x2s = rand(M) * 4.5; 
xt1s = Array([[x1s[i] + 0.5; x2s[i] + 0.5] for i = 1:M])
x1s = rand(M) * 4.5; x2s = rand(M) * 4.5; 
append!(xt1s, Array([[x1s[i] - 5; x2s[i] - 5] for i = 1:M]))

x1s = rand(M) * 4.5; x2s = rand(M) * 4.5; 
xt0s = Array([[x1s[i] + 0.5; x2s[i] - 5] for i = 1:M])
x1s = rand(M) * 4.5; x2s = rand(M) * 4.5; 
append!(xt0s, Array([[x1s[i] - 5; x2s[i] + 0.5] for i = 1:M]))

# Store all the data for later.
xs = [xt1s; xt0s]
ts = [ones(2*M); zeros(2*M)]

# Plot data points.
function plot_data()
    x1 = map(e -> e[1], xt1s)
    y1 = map(e -> e[2], xt1s)
    x2 = map(e -> e[1], xt0s)
    y2 = map(e -> e[2], xt0s)

    Plots.scatter(x1,y1, color="red", clim = (0,1))
    Plots.scatter!(x2, y2, color="blue", clim = (0,1))



Building a Neural Network

The next step is to define a feedforward neural network where we express our parameters as distribtuions, and not single points as with traditional neural networks. The two functions below, unpack and nn_forward are helper functions we need when we specify our model in Turing.

unpack takes a vector of parameters and partitions them between weights and biases. nn_forward constructs a neural network with the variables generated in unpack and returns a prediction based on the weights provided.

The unpack and nn_forward functions are explicity designed to create a neural network with two hidden layers and one output layer, as shown below.


The end of this tutorial provides some code that can be used to generate more general network shapes.

# Turn a vector into a set of weights and biases.
function unpack(nn_params::AbstractVector)
    W₁ = reshape(nn_params[1:6], 3, 2);   
    b₁ = reshape(nn_params[7:9], 3)
    W₂ = reshape(nn_params[10:15], 2, 3); 
    b₂ = reshape(nn_params[16:17], 2)
    Wₒ = reshape(nn_params[18:19], 1, 2); 
    bₒ = reshape(nn_params[20:20], 1)   
    return W₁, b₁, W₂, b₂, Wₒ, bₒ

# Construct a neural network using Flux and return a predicted value.
function nn_forward(xs, nn_params::AbstractVector)
    W₁, b₁, W₂, b₂, Wₒ, bₒ = unpack(nn_params)
    nn = Chain(Dense(W₁, b₁, tanh),
               Dense(W₂, b₂, tanh),
               Dense(Wₒ, bₒ, σ))
    return nn(xs)

The probabalistic model specification below creates a params variable, which has 20 normally distributed variables. Each entry in the params vector represents weights and biases of our neural net.

# Create a regularization term and a Gaussain prior variance term.
alpha = 0.09
sig = sqrt(1.0 / alpha)

# Specify the probabalistic model.
@model bayes_nn(xs, ts) = begin
    # Create the weight and bias vector.
    nn_params ~ MvNormal(zeros(20), sig .* ones(20))
    # Calculate predictions for the inputs given the weights
    # and biases in theta.
    preds = nn_forward(xs, nn_params)
    # Observe each prediction.
    for i = 1:length(ts)
        ts[i] ~ Bernoulli(preds[i])

Inference can now be performed by calling sample. We use the HMC sampler here.

# Perform inference.
N = 5000
ch = sample(bayes_nn(hcat(xs...), ts), HMC(0.05, 4), N);

Now we extract the weights and biases from the sampled chain. We’ll use these primarily in determining how good a classifier our model is.

# Extract all weight and bias parameters.
theta = ch[:nn_params];

Prediction Visualization

We can use MAP estimation to classify our population by using the set of weights that provided the highest log posterior.

# Plot the data we have.

# Find the index that provided the highest log posterior in the chain.
_, i = findmax(ch[:lp]

# Extract the max row value from i.
i = i.I[1]

# Plot the posterior distribution with a contour plot.
x_range = collect(range(-6,stop=6,length=25))
y_range = collect(range(-6,stop=6,length=25))
Z = [nn_forward([x, y], theta[i, :])[1] for x=x_range, y=y_range]
contour!(x_range, y_range, Z)


The contour plot above shows that the MAP method is not too bad at classifying our data.

Now we can visualize our predictions.

$$ p(\tilde{x} | X, \alpha) = \int_{\theta} p(\tilde{x} | \theta) p(\theta | X, \alpha) \approx \sum_{\theta \sim p(\theta | X, \alpha)}f_{\theta}(\tilde{x}) $$

The nn_predict function takes the average predicted value from a network parameterized by weights drawn from the MCMC chain.

# Return the average predicted value across
# multiple weights.
function nn_predict(x, theta, num)
    mean([nn_forward(x, theta[i,:])[1] for i in 1:10:num])

Next, we use the nn_predict function to predict the value at a sample of points where the x and y coordinates range between -6 and 6. As we can see below, we still have a satisfactory fit to our data.

# Plot the average prediction.

n_end = 1500
x_range = collect(range(-6,stop=6,length=25))
y_range = collect(range(-6,stop=6,length=25))
Z = [nn_predict([x, y], theta, n_end)[1] for x=x_range, y=y_range]
contour!(x_range, y_range, Z)


If you are interested in how the predictive power of our Bayesian neural network evolved between samples, the following graph displays an animation of the contour plot generated from the network weights in samples 1 to 1,000.

# Number of iterations to plot.
n_end = 500

anim = @gif for i=1:n_end
    Z = [nn_forward([x, y], theta[i,:])[1] for x=x_range, y=y_range]
    contour!(x_range, y_range, Z, title="Iteration $$i", clim = (0,1))
end every 5

┌ Info: Saved animation to 
│   fn = /home/cameron/code/TuringTutorials/tmp.gif
└ @ Plots /home/cameron/.julia/packages/Plots/cc8wh/src/animation.jl:98