Hierarchical Bayesian Neural Networks with Informative Priors

(c) 2018 by Thomas Wiecki

Imagine you have a machine learning (ML) problem but only small data (gasp, yes, this does exist). This often happens when your data set is nested -- you might have many data points, but only few per category. For example, in ad-tech you may want predict how likely a user will buy a certain product. There could be thousands of products but you only have a small number of measurements (e.g. purchase decisions) per item. Likely, there will be similarities between each product category, but they will all have individual differences too.

Using classic ML methods is unlikely to yield good results in this setting. Firstly, ML is happiest with lots of data. If you want to be fancy and use deep learning, that usually requires even more data to give good performance. Moreover, we can't really make use of the hierarchical structure (product-groupings) we know to be present in our data. So we can then either ignore it and try to learn one model for all categories which does not allow us to capture the differences between categories; or we can fit a ML model to each category separately but do not exploit the similarities between categories. (Technically, you could add the category as a feature and fit a single model but when I tried it on the data presented futher below it was at chance).

Hierarchical Bayesian models work amazingly well in exactly this setting as they allow us to build a model that matches the hierarchical structure present in our data set. Check out my previous blog post The Best Of Both Worlds: Hierarchical Linear Regression in PyMC3 for a refresher.

In this setting we could likely build a hierarchical logistic Bayesian model using PyMC3. However, what if our decision surface is actually more complex and a linear model would not give good performance?

In this blog post I explore how we can take a Bayesian Neural Network (BNN) and turn it into a hierarchical one. Once we built this model we derive an informed prior from it that we can apply back to a simple, non-hierarchical BNN to get the same performance as the hierachical one.

In the ML community, this problem is referred to as multitask transfer learning. Neural networks trained to play multiple Atari games have a hard time because they tend to forget what they learned before fairly quickly when trained on subsequent tasks (catastrophic forgetting), as the old weights get overwritten. Hierarchical BNNs can provide an elegant solution to this problem by sharing the higher-order representations that are useful but still allowing every subnetwork to specialize in a single task.

In [1]:
%matplotlib inline
import numpy as np
from scipy import stats
import matplotlib.pyplot as plt
import seaborn as sns
from warnings import filterwarnings

import pymc3 as pm
from pymc3 import floatX
import theano
import theano.tensor as tt
from pymc3.theanof import set_tt_rng, MRG_RandomStreams

import sklearn
from sklearn import datasets
from sklearn.preprocessing import scale
from sklearn.model_selection import train_test_split
from sklearn.datasets import make_moons

filterwarnings('ignore')
sns.set_style('white')

set_tt_rng(MRG_RandomStreams(42))

cmap = sns.diverging_palette(250, 12, s=85, l=25, as_cmap=True)
cmap_uncertainty = sns.cubehelix_palette(light=1, as_cmap=True)

layer_names = ['w_in_1_grp', 'w_1_2_grp', 'w_2_out_grp']


The data set we are using are our battle tested half-moons as it is simple, non-linear and leads to pretty visualizations. This is what it looks like:

In [2]:
X, Y = make_moons(noise=0.3, n_samples=1000)
plt.scatter(X[Y==0, 0], X[Y==0, 1], label='Class 0')
plt.scatter(X[Y==1, 0], X[Y==1, 1], color='r', label='Class 1')
sns.despine(); plt.legend();


This is just to illustrate what the data generating distribution looks like, we will use way fewer data points, and create different subsets with different rotations.

In [3]:
def rotate(X, deg):
theta = np.radians(deg)
c, s = np.cos(theta), np.sin(theta)
R = np.matrix([[c, -s], [s, c]])

X = X.dot(R)

return np.asarray(X)

In [4]:
np.random.seed(31)

n_samples = 100
n_grps = 18
n_grps_sq = int(np.sqrt(n_grps))
Xs, Ys = [], []
for i in range(n_grps):
# Generate data with 2 classes that are not linearly separable
X, Y = make_moons(noise=0.3, n_samples=n_samples)
X = scale(X)
X = floatX(X)
Y = floatX(Y)

# Rotate the points randomly for each category
rotate_by = np.random.randn() * 90.
X = rotate(X, rotate_by)
Xs.append(X)
Ys.append(Y)

Xs = np.stack(Xs)
Ys = np.stack(Ys)

Xs_train = Xs[:, :n_samples // 2, :]
Xs_test = Xs[:, n_samples // 2:, :]
Ys_train = Ys[:, :n_samples // 2]
Ys_test = Ys[:, n_samples // 2:]

In [5]:
fig, axs = plt.subplots(figsize=(15, 12), nrows=n_grps_sq, ncols=n_grps_sq,
sharex=True, sharey=True)
axs = axs.flatten()
for i, (X, Y, ax) in enumerate(zip(Xs_train, Ys_train, axs)):
ax.scatter(X[Y==0, 0], X[Y==0, 1], label='Class 0')
ax.scatter(X[Y==1, 0], X[Y==1, 1], color='r', label='Class 1')
sns.despine(); ax.legend()
ax.set(title='Category {}'.format(i + 1), xlabel='X1', ylabel='X2')


As you can see, we have 18 categories that share a higher-order structure (the half-moons). However, in the pure data space, no single classifier will be able to do a good job here. Also, because we only have 50 data points in each class, a NN will likely have a hard time producing robust results. But let's actually test this.

Classify each category separately¶

The code for the NN below is explained in my previous blog post on Bayesian Deep Learning.

In [6]:
def construct_flat_nn(ann_input, ann_output):
"""Function to create a flat BNN given data."""
n_hidden = 5

# Initialize random weights between each layer
init_1 = floatX(np.random.randn(X.shape[1], n_hidden))
init_2 = floatX(np.random.randn(n_hidden, n_hidden))
init_out = floatX(np.random.randn(n_hidden))

with pm.Model() as neural_network:
# Weights from input to hidden layer
weights_in_1 = pm.Normal('w_in_1', 0, sd=1,
shape=(X.shape[1], n_hidden),
testval=init_1)

# Weights from 1st to 2nd layer
weights_1_2 = pm.Normal('w_1_2', 0, sd=1,
shape=(n_hidden, n_hidden),
testval=init_2)

# Weights from hidden layer to output
weights_2_out = pm.Normal('w_2_out', 0, sd=1,
shape=(n_hidden,),
testval=init_out)

# Build neural-network using tanh activation function
act_1 = pm.math.tanh(pm.math.dot(ann_input, weights_in_1))
act_2 = pm.math.tanh(pm.math.dot(act_1, weights_1_2))
act_out = pm.math.dot(act_2, weights_2_out)

# Binary classification -> Bernoulli likelihood
out = pm.Bernoulli('out',
logit_p=act_out,
observed=ann_output)
return neural_network


Next, we have our function to create the BNN, sample from it, and generate predicitions. You can skip this one.

In [98]:
def fit_and_eval_bnn(X_train, X_test, Y_train, Y_test, grid, dummy_out, bnn_func, bnn_kwargs=None, sample_kwargs=None):
"""Utility function to create a BNN from a function, sample from it, and create predictions."""
if bnn_kwargs is None:
bnn_kwargs = {}

if sample_kwargs is None:
sample_kwargs = {'chains': 1, 'progressbar': False}

ann_input = theano.shared(X_train)
ann_output = theano.shared(Y_train)
model = bnn_func(ann_input, ann_output, **bnn_kwargs)

with model:
# fit model
trace = pm.sample(**sample_kwargs)
# sample posterior predictive
ppc_train = pm.sample_ppc(trace, samples=500, progressbar=False)

# Use probability of > 0.5 to assume prediction of class 1
pred_train = ppc_train['out'].mean(axis=0) > 0.5

# Make predictions on test-set
ann_input.set_value(X_test)
ann_output.set_value(Y_test)
ppc_test = pm.sample_ppc(trace, samples=500, progressbar=False)

pred_test = ppc_test['out'].mean(axis=0) > 0.5

# Evaluate classifier over grid
ann_input.set_value(grid)
ann_output.set_value(dummy_out)
ppc_grid = pm.sample_ppc(trace, samples=500,
progressbar=False)['out']

return pred_train, pred_test, ppc_grid, trace


Next, we loop over each category and fit a different BNN to each one. Each BNN has its own weights and there is no connection between them. But note that because we are Bayesians, we place priors on our weights. In this case these are standard normal priors that act as regularizers to keep our weights close to zero.

We use NUTS sampling here because the hierarchical model further below has a more complex posterior (see Why hierarchical models are awesome, tricky, and Bayesian for an explanation) and I wanted the results to be comparable. All simulations in here work with ADVI as well but the results don't look quite as strong.

In [ ]:
Ys_pred_train = []
Ys_pred_test = []
grid_eval = []

grid = pm.floatX(np.mgrid[-3:3:100j,-3:3:100j])
grid_2d = grid.reshape(2, -1).T
dummy_out = np.ones(grid.shape[1], dtype=np.int8)

for X_train, Y_train, X_test, Y_test in zip(Xs_train, Ys_train, Xs_test, Ys_test):
pred_train, pred_test, ppc_grid, trace_flat = \
fit_and_eval_bnn(X_train, X_test,
Y_train, Y_test,
grid_2d, dummy_out,
construct_flat_nn)
Ys_pred_train.append(pred_train)
Ys_pred_test.append(pred_test)
grid_eval.append(ppc_grid)

Ys_pred_train = np.stack(Ys_pred_train)
Ys_pred_test = np.stack(Ys_pred_test)
ppc_grid_single = np.stack(grid_eval)

In [9]:
print ("Train accuracy = {:.2f}%".format(100*np.mean(Ys_pred_train == Ys_train)))

Train accuracy = 86.78%

In [10]:
print ("Test accuracy = {:.2f}%".format(100*np.mean(Ys_pred_test == Ys_test)))

Test accuracy = 83.78%


OK, that doesn't seem so bad. Now let's look at the decision surfaces -- i.e. what the classifier thinks about each point in the data space.

In [11]:
fig, axs = plt.subplots(figsize=(15, 12), nrows=n_grps_sq, ncols=n_grps_sq, sharex=True, sharey=True)
axs = axs.flatten()
for i, (X, Y_pred, Y_true, ax) in enumerate(zip(Xs_train, Ys_pred_train, Ys_train, axs)):
contour = ax.contourf(grid[0], grid[1], ppc_grid_single[i, ...].mean(axis=0).reshape(100, 100), cmap=cmap)
ax.scatter(X[Y_true == 0, 0], X[Y_true == 0, 1], label='Class 0')
ax.scatter(X[Y_true == 1, 0], X[Y_true == 1, 1], color='r', label='Class 1')
sns.despine(); ax.legend()


That doens't look all that convincing. We know from the data generation process as well as from the previous blog post Bayesian Deep learning using the same data set that it give us a "Z"-shaped decision surface. So what happens is we don't have enough data to properly estimate the non-linearity in every category.

Hierarchical Bayesian Neural Network¶

Can we do better? You bet!

It's actually quite straight-forward to turn this into one big hierarchical model for all categories, rather than many individual ones. Let's call the weight connecting neuron $i$ in layer 1 to neuron $j$ in layer 2 in category $c$ $w_{i, j, c}$ (I just omit the layer index for simplicity in notation). Rather than placing a fixed prior as we did above (i.e. $w_{i, j, c} \sim \mathcal{N}(0, 1^2)$), we will assume that each weight comes from an overarching group distribution: $w_{i, j, c} \sim \mathcal{N}(\mu_{i, j}, \sigma^2)$. The key is that we will estimate $\mu_{i, j}$ and $\sigma$ simultaneously from data.

Why not allow for different $\sigma_{i,j}^2$ per connection you might ask? Mainly just to make our life simpler and because it works well enough.

Note that we create a very rich model here. Every individual weight has its own hierarchical structure with a single group mean parameter and 16 per-category weights distributed around the group mean. While this creates a big amount of group distributions (as many as the flat NN had weights) there is no problem with this per-se, although it might be a bit unusual. One might argue that this model is quite complex and while that's true, in terms of degrees-of-freedom, this model is simpler than the unpooled one above (more on this below).

As for the code, we stack weights along a 3rd dimenson to get separate weights for each group. That way, through the power of broadcasting, the linear algebra works out almost the same as before.

In [12]:
def construct_hierarchical_nn(Xs, Ys):
n_hidden = 5
n_grps = Xs.shape[0].eval()
n_data = Xs.shape[2].eval()
# Initialize random weights between each layer
init_1 = floatX(np.random.randn(n_data, n_hidden))
init_2 = floatX(np.random.randn(n_hidden, n_hidden))
init_out = floatX(np.random.randn(n_hidden))

with pm.Model() as neural_network:
# Group mean distribution for input to hidden layer
weights_in_1_grp = pm.Normal('w_in_1_grp', 0, sd=1,
shape=(n_data, n_hidden),
testval=init_1)
# Group standard-deviation
weights_in_1_grp_sd = pm.HalfNormal('w_in_1_grp_sd', sd=1.)

# Group mean distribution for weights from 1st to 2nd layer
weights_1_2_grp = pm.Normal('w_1_2_grp', 0, sd=1,
shape=(n_hidden, n_hidden),
testval=init_2)
weights_1_2_grp_sd = pm.HalfNormal('w_1_2_grp_sd', sd=1.)

# Group mean distribution from hidden layer to output
weights_2_out_grp = pm.Normal('w_2_out_grp', 0, sd=1,
shape=(n_hidden,),
testval=init_out)
weights_2_out_grp_sd = pm.HalfNormal('w_2_out_grp_sd', sd=1.)

# Separate weights for each different model, just add a 3rd dimension
# of weights
weights_in_1_raw = pm.Normal('w_in_1',
shape=(n_grps, n_data, n_hidden))
# Non-centered specification of hierarchical model
# see https://twiecki.github.io/blog/2017/02/08/bayesian-hierchical-non-centered/
weights_in_1 = weights_in_1_raw * weights_in_1_grp_sd + weights_in_1_grp

weights_1_2_raw = pm.Normal('w_1_2',
shape=(n_grps, n_hidden, n_hidden))
weights_1_2 = weights_1_2_raw * weights_1_2_grp_sd + weights_1_2_grp

weights_2_out_raw = pm.Normal('w_2_out',
shape=(n_grps, n_hidden))

weights_2_out = weights_2_out_raw * weights_2_out_grp_sd + weights_2_out_grp

# Build neural-network using tanh activation function
# tt.batched_dot just calls .dot along an axis
act_1 = pm.math.tanh(tt.batched_dot(Xs, weights_in_1))
act_2 = pm.math.tanh(tt.batched_dot(act_1, weights_1_2))
act_out = tt.batched_dot(act_2, weights_2_out)

# Binary classification -> Bernoulli likelihood
out = pm.Bernoulli('out', logit_p=act_out, observed=Ys)

return neural_network

In [68]:
sample_kwargs = dict(init='advi+adapt_diag', tune=5000, chains=1,
nuts_kwargs={'target_accept': 0.9},
progressbar=False)
grid_3d = np.repeat(grid_2d[None, ...], n_grps, axis=0)
dummy_out_3d = np.ones((n_grps, grid.shape[1]), dtype=np.int8)

Ys_hierarchical_pred_train, Ys_hierarchical_pred_test, ppc_grid, trace_hier = \
fit_and_eval_bnn(
Xs_train, Xs_test,
Ys_train, Ys_test,
grid_3d, dummy_out_3d,
construct_hierarchical_nn,
sample_kwargs=sample_kwargs)

Auto-assigning NUTS sampler...
Initializing NUTS using advi+adapt_diag...
Convergence achieved at 54100
Interrupted at 54,099 [27%]: Average Loss = 522.66
Sequential sampling (1 chains in 1 job)
NUTS: [w_2_out, w_1_2, w_in_1, w_2_out_grp_sd, w_2_out_grp, w_1_2_grp_sd, w_1_2_grp, w_in_1_grp_sd, w_in_1_grp]
Only one chain was sampled, this makes it impossible to run some convergence checks

In [69]:
print('Train accuracy = {:.2f}%'.format(100*np.mean(Ys_hierarchical_pred_train == Ys_train)))

Train accuracy = 91.11%

In [70]:
print('Test accuracy = {:.2f}%'.format(100*np.mean(Ys_hierarchical_pred_test == Ys_test)))

Test accuracy = 89.78%


Great -- we get higher train and test accuracy. Let's look at what the classifier has learned for each category.

In [88]:
fig, axs = plt.subplots(figsize=(15, 12), nrows=n_grps_sq, ncols=n_grps_sq, sharex=True, sharey=True)
axs = axs.flatten()
for i, (X, Y_pred, Y_true, ax) in enumerate(zip(Xs_train, Ys_hierarchical_pred_train, Ys_train, axs)):
contour = ax.contourf(grid[0], grid[1], ppc_grid[:, i, :].mean(axis=0).reshape(100, 100), cmap=cmap)
ax.scatter(X[Y_true == 0, 0], X[Y_true == 0, 1], label='Class 0')
ax.scatter(X[Y_true == 1, 0], X[Y_true == 1, 1], color='r', label='Class 1')
sns.despine(); ax.legend()


Awesome! By (partially) pooling the data for each individual category we actually manage to retrieve the non-linearity. This is the strength of hierarchical models: we model the similarities of individual categories and their differences, sharing statistical power to the degree it's useful.

Of course, as we are in a Bayesian framework we also get the uncertainty estimate in our predictions:

In [89]:
fig, axs = plt.subplots(figsize=(15, 12), nrows=n_grps_sq, ncols=n_grps_sq, sharex=True, sharey=True)
axs = axs.flatten()
for i, (X, Y_pred, Y_true, ax) in enumerate(zip(Xs_train, Ys_hierarchical_pred_train, Ys_train, axs)):
contour = ax.contourf(grid[0], grid[1], ppc_grid[:, i, :].std(axis=0).reshape(100, 100),
cmap=cmap_uncertainty)
ax.scatter(X[Y_true == 0, 0], X[Y_true == 0, 1], label='Class 0')
ax.scatter(X[Y_true == 1, 0], X[Y_true == 1, 1], color='r', label='Class 1')
sns.despine(); ax.legend()


Further analysis¶

There are a couple of things we might ask at this point. For example, how much does each layer specialize its weight per category. To answer this we can look at the group standard-deviation which informs us how much each weight is allowed to deviate from its group mean.

In [90]:
pm.traceplot(trace_hier, varnames=['w_in_1_grp_sd', 'w_1_2_grp_sd', 'w_2_out_grp_sd']);


Interestingly, it seems that the specialization of the individual category sub-models is happening at the last layer where weights change most strongly from their group mean (as group variance is highest). I had assumed that this would happen at the first layer, based on what I found in my earlier blog post Random-Walk Bayesian Deep Networks: Dealing with Non-Stationary Data where the first layer acted as a rotation-layer.

Another interesting property of hierarchical models reveals itself here. As the group standard deviation is small for the weights in layers 1 and 2, it means these weights will be close to their group mean, reducing the effective number of parameters (degrees of freedom) of this model. This is different from the separate models where no similarities could be exploited to reduce the effective number of parameters. So from this perspective, the hierarchical model is simpler than the sum of the separate ones above.

Finally, I wondered what the group-model actually learned. To get at that, we can use the trace of $\mu_{i,j}$ from the hierarchical model and pass it to the non-hierarchical model as if we trained these weights directly on a single data set.

In [91]:
trace_flat._straces[0].samples['w_in_1'] = trace_hier['w_in_1_grp']
trace_flat._straces[0].samples['w_1_2'] = trace_hier['w_1_2_grp']
trace_flat._straces[0].samples['w_2'] = trace_hier['w_2_out_grp']

ann_input = theano.shared(X_train)
ann_output = theano.shared(Y_train)
neural_network = construct_flat_nn(ann_input, ann_output)

with neural_network:
# Evaluate classifier over grid
ann_input.set_value(grid_2d)
ann_output.set_value(dummy_out)
ppc_grid_hier2 = pm.sample_ppc(trace_flat, samples=500,
progressbar=False)['out']

In [106]:
contour = plt.contourf(grid[0], grid[1], ppc_grid_hier2.mean(axis=0).reshape(100, 100), cmap=cmap)
plt.xlabel('X1'); plt.ylabel('X2'); plt.title('Decision surface of the group weights');