# Blog Archives

## Introduction

Artificial neural networks (ANNs) are a powerful class of models used for nonlinear regression and classification tasks that are motivated by biological neural computation. The general idea behind ANNs is pretty straightforward: map some input onto a desired target value using a distributed cascade of nonlinear transformations (see Figure 1). However, for many, myself included, the learning algorithm used to train ANNs can be difficult to get your head around at first. In this post I give a step-by-step walk-through of the derivation of gradient descent learning algorithm commonly used to train ANNs (aka the backpropagation algorithm) and try to provide some high-level insights into the computations being performed during learning.

Figure 1: Diagram of an artificial neural network with one hidden layer

### Some Background and Notation

An ANN consists of an input layer, an output layer, and any number (including zero) of hidden layers situated between the input and output layers. Figure 1 diagrams an ANN with a single hidden layer. The feed-forward computations performed by the ANN are as follows: The signals from the input layer $a_i$ are multiplied by a set of fully-connected weights $w_{ij}$ connecting the input layer to the hidden layer. These weighted signals are then summed and combined with a bias $b_i$ (not displayed in the graphical model in Figure 1). This calculation forms the pre-activation signal $z_j = b_j + \sum_i a_i w_{ij}$ for the hidden layer. The pre-activation signal is then transformed by the hidden layer activation function $g_j$ to form the feed-forward activation signals leaving leaving the hidden layer $a_j$. In a similar fashion, the hidden layer activation signals $a_j$ are multiplied by the weights connecting the hidden layer to the output layer $w_{jk}$, a bias $b_k$ is added, and the resulting signal is transformed by the output activation function $g_k$ to form the network output $a_k$. The output is then compared to a desired target $t_k$ and the error between the two is calculated.

Training a neural network involves determining the set of parameters $\theta = \{\mathbf{W},\mathbf{b}\}$ that minimize the errors that the network makes. Often the choice for the error function is the sum of the squared difference between the target values $t_k$ and the network output $a_k$ (for more detail on this choice of error function see):

$\Large{\begin{array}{rcl} E &=& \frac{1}{2} \sum_{k \in K}(a_k - t_k)^2 \end{array}}$

Equation (1)

This problem can be solved using gradient descent, which requires determining $\frac{\partial E}{\partial \theta}$ for all $\theta$ in the model. Note that, in general, there are two sets of parameters: those parameters that are associated with the output layer (i.e. $\theta_k = \{w_{jk}, b_k\}$), and thus directly affect the network output error; and the remaining parameters that are associated with the hidden layer(s), and thus affect the output error indirectly.

Before we begin, let’s define the notation that will be used in remainder of the derivation. Please refer to Figure 1 for any clarification.

• ${z_j}$: input to node $j$ for layer $l$
• ${g_j}$: activation function for node $j$ in layer $l$ (applied to ${z_j}$)
• $a_j=g_j(z_j)$: ouput/activation of node $j$ in layer $l$
• ${w_{ij}}$: weights connecting node $i$ in layer $(l-1)$ to node $j$ in layer $l$
• ${b_{j}}$: bias for unit $j$ in layer $l$
• ${t_{k}}$: target value for node $k$ in the output layer

## Gradients for Output Layer Weights

### Output layer connection weights, $w_{jk}$

Since the output layer parameters directly affect the value of the error function, determining the gradients for those parameters is fairly straight-forward:

$\Large{\begin{array}{rcl} \frac{\partial E }{\partial w_{jk}} &=& \frac{1}{2} \sum_{k \in K}(a_k - t_k)^2 \\ &=& (a_k - t_k)\frac{\partial}{\partial w_{jk}}(a_k - t_k) \end{array}}$

Equation (2)

Here, we’ve used the Chain Rule. (Also notice that the summation disappears in the derivative. This is because when we take the partial derivative with respect to the $j$-th dimension/node, the only term that survives in the error gradient is $j$-th, and thus we can ignore the remaining terms in the summation). The derivative with respect to $t_k$ is zero because it does not depend on $w_{jk}$. Also, we note that $a_k = g(z_k)$. Thus

$\Large{\begin{array}{rcl}\frac{\partial E }{\partial w_{jk}} &=& (a_k - t_k)\frac{\partial}{\partial w_{jk}}a_k \\ &=& (a_k - t_k)\frac{\partial}{\partial w_{jk}}g_k(z_k) \\ &=& (a_k - t_k)g_k'(z_k)\frac{\partial}{\partial w_{jk}}z_k, \end{array}}$

Equation (3)

where, again we use the Chain Rule. Now, recall that $z_k = b_j + \sum_j g_j(z_j)w_{jk}$ and thus $\frac{\partial z_{k}}{\partial w_{jk}} = g_j(z_j) = a_j$, giving:

$\Large{\begin{array}{rcl} \frac{\partial E }{\partial w_{jk}} &=& (a_k - t_k)g_k'(z_k)a_j \end{array}}$

Equation (4)

The gradient of the error function with respect to the output layer weights is a product of three terms. The first term is the difference between the network output and the target value $t_k$. The second term is the derivative of output layer activation function. And the third term is the activation output of node j in the hidden layer.

If we define $\delta_k$ to be all the terms that involve index k:

$\Large{\begin{array}{rcl} \delta_k &=& (a_k - t_k)g_k'(z_k)\end{array}}$

we obtain the following expression for the derivative of the error with respect to the output weights $w_{jk}$:

$\Large{\begin{array}{rcl} \frac{\partial E }{\partial w_{jk}} = \delta_k a_j \end{array}}$

Equation (5)

Here the $\delta_k$ terms can be interpreted as the network output error after being back-propagated through the output activation function, thus creating an error “signal”. Loosely speaking, Equation (5) can be interpreted as determining how much each $w_{jk}$ contributes to the error signal by weighting the error signal by the magnitude of the output activation from the previous (hidden) layer associated with each weight (see Figure 1). The gradients with respect to each parameter are thus considered to be the “contribution” of the parameter to the error signal and should be negated during learning. Thus the output weights are updated as $w_{jk}\leftarrow w_{jk} - \eta \frac{\partial E }{\partial w_{jk}}$, where $\eta$ is some step size (“learning rate”) along the negative gradient.

As we’ll see shortly, the process of backpropagating the error signal can iterate all the way back to the input layer by successively projecting $\delta_k$ back through $w_{jk}$, then through the activation function for the hidden layer via $g'_j$ to give the error signal $\delta_j$, and so on. This backpropagation concept is central to training neural networks with more than one layer.

### Output layer biases, $\Large{b_{k}}$

As far as the gradient with respect to the output layer biases, we follow the same routine as above for $w_{jk}$. However, the third term in Equation (3) is $\frac{\partial}{\partial b_k} z_k = \frac{\partial}{\partial b_k} \left[ b_k + \sum_j g_j(z_j)\right] = 1$, giving the following gradient for the output biases:

$\Large{\begin{array}{rcl} \frac{\partial E }{\partial b_k} &=& (a_k - t_k)g_k'(z_k)(1) \\ &=& \delta_k \end{array}}$

Equation (6)

Thus the gradient for the biases is simply the back-propagated error from the output units. One interpretation of this is that the biases are weights on activations that are always equal to one, regardless of the feed-forward signal. Thus the bias gradients aren’t affected by the feed-forward signal, only by the error.

## Gradients for Hidden Layer Weights

Due to the indirect affect of the hidden layer on the output error, calculating the gradients for the hidden layer weights $w_{ij}$  is somewhat more involved. However, the process starts just the same:

$\Large{\begin{array}{rcl} \frac{\partial E }{\partial w_{ij}}&=&\frac{1}{2} \sum_{k \in K}(a_k - t_k)^2 \\ &=& \sum_{k \in K} (a_k - t_k) \frac{\partial}{\partial w_{ij}}a_k \end{array}}$

Notice here that the sum does not disappear because, due to the fact that the layers are fully connected, each of the hidden unit outputs affects the state of each output unit. Continuing on, noting that $a_k = g_k(z_k)$

$\Large{\begin{array}{rcl} \frac{\partial E }{\partial w_{ij}}&=& \sum_{k \in K} (a_k - t_k) \frac{\partial }{\partial w_{ij}}g_k(z_k) \\ &=& \sum_{k \in K} (a_k - t_k)g'_k(z_k)\frac{\partial }{\partial w_{ij}}z_k \end{array}}$

Equation (7)

Here, again we use the Chain Rule. Ok, now here’s where things get “slightly more involved”. Notice that the partial derivative in the third term in Equation (7) is with respect to $w_{ij}$, but the target $z_j$ is a function of index $j$. How the heck do we deal with that!? Well, if we expand $z_k$, we find that it is composed of other sub functions (also see Figure 1):

$\Large{\begin{array}{rcl} z_k &=& b_k + \sum_j a_jw_{jk} \\ &=& b_k + \sum_j g_j(z_j)w_{jk} \\ &=& b_k + \sum_j g_j(b_i + \sum_i z_i w_{ij})w_{jk}\end{array}}$

Equation (8)

From the last term in Equation (8) we see that $z_k$ is indirectly dependent on $w_{ij}$.  Equation (8) also suggests that we can use the Chain Rule to calculate $\frac{\partial z_k }{\partial w_{ij}}$. This is probably the trickiest part of the derivation, and goes like…

$\Large{\begin{array}{rcl} \frac{\partial z_k }{\partial w_{ij}} &=& \frac{\partial z_k}{\partial a_j}\frac{\partial a_j}{\partial w_{ij}} \\ &=& \frac{\partial}{\partial a_j}a_jw_{jk}\frac{\partial a_j}{\partial w_{ij}} \\ &=& w_{jk}\frac{\partial a_j}{\partial w_{ij}} \\ &=& w_{jk}\frac{\partial g_j(z_j)}{\partial w_{ij}} \\ &=& w_{jk}g_j'(z_j)\frac{\partial z_j}{\partial w_{ij}} \\ &=& w_{jk}g_j'(z_j)\frac{\partial}{\partial w_{ij}}(b_i + \sum_i a_i w_{ij}) \\ &=& w_{jk}g_j'(z_j)a_i \end{array}}$

Equation (9)

Now, plugging Equation (9) into $z_k$ in Equation (7) gives the following for $\frac{\partial E}{\partial w_{ij}}$:

$\Large{\begin{array}{rcl} \frac{\partial E }{\partial w_{ij}}&=& \sum_{k \in K} (a_k - t_k)g'_k(z_k)w_{jk} g'_j(z_j)a_i \\ &=& g'_j(z_j)a_i \sum_{k \in K} (a_k - t_k)g'_k(z_k)w_{jk} \\ &=& a_i g'_j(z_j) \sum_{k \in K} \delta_k w_{jk} \end{array}}$

Equation (10)

Notice that the gradient for the hidden layer weights has a similar form to that of the gradient for the output layer weights. Namely the gradient is some term weighted by the output activations from the layer below ($a_i$). For the output weight gradients, the term that was weighted by $a_j$ was the back-propagated error signal $\delta_k$ (i.e. Equation (5)). Here, the weighted term includes $\delta_k$, but the error signal is further projected onto $w_{jk}$ and then weighted by the derivative of hidden layer activation function $g'_j$. Thus, the gradient for the hidden layer weights is simply the output error signal backpropagated to the hidden layer, then weighted by the input to the hidden layer. To make this idea more explicit, we can define the resulting error signal backpropagated to layer $j$ as $\delta_j$, and includes all terms in Equation (10) that involve index $j$. This definition results in the following gradient for the hidden unit weights:

$\Large{\begin{array}{rcl} \frac{\partial E }{\partial w_{ij}}&=& a_i g'_j(z_j) \sum_{k \in K} \delta_k w_{jk} \\ &=& \delta_j a_i \\ \text{where} \\ \delta_j &=& g'_j(z_j) \sum_{k \in K} \delta_k w_{jk} \end{array}}$

Equation (11)

This suggests that in order to calculate the weight gradients at any layer $l$ in an arbitrarily-deep neural network, we simply need to calculate the backpropagated error signal that reaches that layer $\delta_l$ and weight it by the feed-forward signal $a_{l-1}$feeding into that layer! Analogously, the gradient for the hidden layer weights can be interpreted as a proxy for the “contribution” of the weights to the output error signal, which can only be observed–from the point of view of the weights–by backpropagating the error signal to the hidden layer.

### Output layer biases, $\Large{w_{ij}}$

Calculating the gradients for the hidden layer biases follows a very similar procedure to that for the hidden layer weights where, as in Equation (9), we use the Chain Rule to calculate $\frac{\partial z_k}{\partial b_i}$. However, unlike Equation (9) the third term that results for the biases is slightly different:

$\Large{\begin{array}{rcl} \frac{\partial z_k }{\partial b_i} &=& w_{jk}g_j'(z_j)\frac{\partial z_j}{\partial b_i} \\ &=& w_{jk}g_j'(z_j)\frac{\partial}{\partial b_i}(b_i + \sum_i a_i w_{ij}) \\ &=& w_{jk}g_j'(z_j)(1), \\ \text{giving} \\ \frac{\partial E }{\partial b_i}&=& g'_j(z_j) \sum_{k \in K} \delta_k w_{jk} \\ &=& \delta_j \end{array}}$

Equation (12)

In a similar fashion to calculation of the bias gradients for the output layer, the gradients for the hidden layer biases are simply the backpropagated error signal reaching that layer. This suggests that we can also calculate the bias gradients at any layer $l$ in an arbitrarily-deep network by simply calculating the backpropagated error signal reaching that layer $\delta_l$!

## Wrapping up

In this post we went over some of the formal details of the backpropagation learning algorithm. The math covered in this post allows us to train arbitrarily deep neural networks by re-applying the same basic computations. Those computations are:

1. Calculated the feed-forward signals from the input to the output.
2. Calculate output error $E$ based on the predictions $a_k$ and the target $t_k$
3. Backpropagate the error signals by weighting it by the weights in previous layers and the gradients of the associated activation functions
4. Calculating the gradients $\frac{\partial E}{\partial \theta}$ for the parameters based on the backpropagated error signal and the feedforward signals from the inputs.
5. Update the parameters using the calculated gradients $\theta \leftarrow \theta - \eta\frac{\partial E}{\partial \theta}$

The only real constraints on model construction is ensuring that the error function $E$ and the activation functions $g_l$ are differentiable. For more details on implementing ANNs and seeing them at work, stay tuned for the next post.

## Inverse Transform Sampling

There are a number of sampling methods used in machine learning, each of which has various strengths and/or weaknesses depending on the nature of the sampling task at hand. One simple method for generating samples from distributions with closed-form descriptions is Inverse Transform (IT) Sampling.

The idea behind IT Sampling is that the probability mass for a random variable $X$ distributed according to the probability density function $f(x)$ integrates to one and therefore the cumulative distribution function $C(x)$ can be used to map from values in the interval $(0,1)$ (i.e. probabilities) to the domain of $f(x)$. Because it is easy to sample values $z$ uniformly from the interval $(0,1)$, we can use the inverse of the CDF $C(x)^{-1}$ to transform these sampled probabilities into samples $x$. The code below demonstrates this process at work in order to sample from a student’s t distribution with 10 degrees of freedom.

rand('seed',12345)

% DEGREES OF FREEDOM
dF = 10;
x = -3:.1:3;
Cx = cdf('t',x,dF)
z = rand;

% COMPARE VALUES OF
zIdx = min(find(Cx>z));

% DRAW SAMPLE
sample = x(zIdx);

% DISPLAY
figure; hold on
plot(x,Cx,'k','Linewidth',2);
plot([x(1),x(zIdx)],[Cx(zIdx),Cx(zIdx)],'r','LineWidth',2);
plot([x(zIdx),x(zIdx)],[Cx(zIdx),0],'b','LineWidth',2);
plot(x(zIdx),z,'ko','LineWidth',2);
text(x(1)+.1,z + .05,'z','Color','r')
text(x(zIdx)+.05,.05,'x_{sampled}','Color','b')
ylabel('C(x)')
xlabel('x')
hold off


IT Sampling from student’s-t(10)

However, the scheme used to create to plot above is inefficient in that one must compare current values of $z$ with the $C(x)$ for all values of $x$. A much more efficient method is to evaluate $C^{-1}$ directly:

1. Derive $C^{-1}(x)$ (or a good approximation) from $f(x)$
2. for $i = 1:n$
• – draw $z_i$ from $Unif(0,1)$
• $x_i = CDF^{-1}(z_i)$
• – end for

The IT sampling process is demonstrated in the next chunk of code to sample from the Beta distribution, a distribution for which $C^{-1}$  is easy to approximate using Netwon’s method (which we let MATLAB do for us within the function icdf.m)

rand('seed',12345)
nSamples = 1000;

% BETA PARAMETERS
alpha = 2; beta = 10;

% DRAW PROPOSAL SAMPLES
z = rand(1,nSamples);

% EVALUATE PROPOSAL SAMPLES AT INVERSE CDF
samples = icdf('beta',z,alpha,beta);
bins = linspace(0,1,50);
counts = histc(samples,bins);
probSampled = counts/sum(counts)
probTheory = betapdf(bins,alpha,beta);

% DISPLAY
b = bar(bins,probSampled,'FaceColor',[.9 .9 .9]);
hold on;
t = plot(bins,probTheory/sum(probTheory),'r','LineWidth',2);
xlim([0 1])
xlabel('x')
ylabel('p(x)')
legend([t,b],{'Theory','IT Samples'})
hold off


Inverse Transform Sampling of Beta(2,10)

## Wrapping Up

The IT sampling method is generally only used for univariate distributions where $C^{-1}$ can be computed in closed form, or approximated. However, it is a nice example of how uniform random variables can be used to sample from much more complicated distributions.