# Neural Networks: The mechanics of backpropagation

The initial work in the  ‘Backpropagation Algorithm’  started in the 1980’s and led to an explosion of interest in Neural Networks and  the application of backpropagation

The ‘Backpropagation’ algorithm computes the minimum of an error function with respect to the weights in the Neural Network. It uses the method of gradient descent. The combination of weights in a multi-layered neural network, which minimizes the error/cost function is considered to be a solution of the learning problem.

In the Neural Network above
$out_{o1} =\sum_{i} w_{i}*x_{i}$
$E = 1/2(target - out)^{2}$
$\partial E/\partial out= 1/2*2*(target - out) *-1 = -(target - out)$
$\partial E/\partial w_{i} =\partial E/\partial y* \partial y/\partial w_{i}$
$\partial E/\partial w_{i} = -(target - out) * x_{i}$

Checkout my book ‘Deep Learning from first principles: Second Edition – In vectorized Python, R and Octave’. My book starts with the implementation of a simple 2-layer Neural Network and works its way to a generic L-Layer Deep Learning Network, with all the bells and whistles. The derivations have been discussed in detail. The code has been extensively commented and included in its entirety in the Appendix sections. My book is available on Amazon as paperback ($18.99) and in kindle version($9.99/Rs449).

Perceptrons and single layered neural networks can only classify, if the sample space is linearly separable. For non-linear decision boundaries, a multi layered neural network with  backpropagation is required to generate more complex boundaries.The backpropagation algorithm, computes the minimum of the error function in weight space using the method of gradient descent. This computation of the gradient, requires the activation function to be both differentiable and continuous. Hence the sigmoid or logistic function is typically chosen as the activation function at every layer.

This post looks at a 3 layer neural network with 1 input, 1 hidden and 1 output. To a large extent this post is based on Matt Mazur’s detailed “A step by step backpropagation example“, and Prof Hinton’s “Neural Networks for Machine Learning” at Coursera and a few other sources.

While Matt Mazur’s post uses example values, I generate the formulas for the gradient derivatives for each weight in the hidden and input layers. I intend to implement a vector version of backpropagation in Octave, R and Python. So this post is a prequel to that.

The 3 layer neural network is as below

Some basic derivations which are used in backpropagation

Chain rule of differentiation
Let y=f(u)
and u=g(x) then
$\partial y/\partial x = \partial y/\partial u * \partial u/\partial x$

An important result
$y=1/(1+e^{-z})$
Let $x= 1 + e^{-z}$  then
$y = 1/x$
$\partial y/\partial x = -1/x^{2}$
$\partial x/\partial z = -e^{-z}$

Using the chain rule of differentiation we get
$\partial y/\partial z = \partial y/\partial x * \partial x/\partial z$
$=-1/(1+e^{-z})^{2}* -e^{-z} = e^{-z}/(1+e^{-z})^{2}$
Therefore $\partial y/\partial z = y(1-y)$                                   -(A)

1) Feed forward network
The net output at the 1st hidden layer
$in_{h1} = w_{1}i_{1} + w_{2}i_{2} + b_{1}$
$in_{h2} = w_{3}i_{1} + w_{4}i_{2} + b_{1}$

The sigmoid/logistic function function is used to generate the activation outputs for each hidden layer. The sigmoid is chosen because it is continuous and also has a continuous derivative

$out_{h1} = 1/1+e^{-in_{h1}}$
$out_{h2} = 1/1+e^{-in_{h2}}$

The net output at the output layer
$in_{o1} = w_{5}out_{h_{1}} + w_{6}out_{h_{2}} + b_{2}$
$in_{o2} = w_{7}out_{h_{1}} + w_{8}out_{h_{2}} + b_{2}$

Total error
$E_{total} = 1/2\sum (target - output)^{2}$
$E_{total} = E_{o1} + E_{o2}$
$E_{total} = 1/2(target_{o_{1}} - out_{o_{1}})^{2} + 1/2(target_{o_{2}} - out_{o_{2}})^{2}$

2)The backwards pass
In the backward pass we need to compute how the squared error changes with changing weight. i.e we compute $\partial E_{total}/\partial w_{i}$ for each weight $w_{i}$. This is shown below

A squared error is assumed

Error gradient  with $w_{5}$

$\partial E_{total}/\partial w_{5} = \partial E_{total}/\partial out_{o_{1}} * \partial out_{o_{1}}/\partial in_{o_{1}} * \partial in_{o_{1}}/ \partial w_{5}$                -(B)

Since
$E_{total} = 1/2\sum (target - output)^{2}$
$E_{total} = 1/2(target_{o_{1}} - out_{o_{1}})^{2} + 1/2(target_{o_{2}} - out_{o_{2}})^{2}$
$\partial E _{total}/\partial out_{o1} = \partial E_{o1}/\partial out_{o1} + \partial E_{o2}/\partial out_{o1}$
$\partial E _{total}/\partial out_{o1} = \partial /\partial _{out_{o1}}[1/2(target_{01}-out_{01})^{2}- 1/2(target_{02}-out_{02})^{2}]$
$\partial E _{total}/\partial out_{o1} = 2 * 1/2*(target_{01} - out_{01}) *-1 + 0$

Now considering the 2nd term in (B)
$\partial out_{o1}/\partial in_{o1} = \partial/\partial in_{o1} [1/(1+e^{-in_{o1}})]$

Using result (A)
$\partial out_{o1}/\partial in_{o1} = \partial/\partial in_{o1} [1/(1+e^{-in_{o1}})] = out_{o1}(1-out_{o1})$

The 3rd term in (B)
$\partial in_{o1}/\partial w_{5} = \partial/\partial w_{5} [w_{5}*out_{h1} + w_{6}*out_{h2}] = out_{h1}$
$\partial E_{total}/\partial w_{5}=-(target_{o1} - out_{o1}) * out_{o1} *(1-out_{o1}) * out_{h1}$

Having computed $\partial E_{total}/\partial w_{5}$, we now perform gradient descent, by computing a new weight, assuming a learning rate $\alpha$
$w_{5}^{+} = w_{5} - \alpha * \partial E_{total}/\partial w_{5}$

If we do this for $\partial E_{total}/\partial w_{6}$ we would get
$\partial E_{total}/\partial w_{6}=-(target_{02} - out_{02}) * out_{02} *(1-out_{02}) * out_{h2}$

3)Hidden layer

We now compute how the total error changes for a change in weight $w_{1}$
$\partial E_{total}/\partial w_{1}= \partial E_{total}/\partial out_{h1}* \partial out_{h1}/\partial in_{h1} * \partial in_{h1}/\partial w_{1}$ – (C)

Using
$E_{total} = E_{o1} + E_{o2}$ we get
$\partial E_{total}/\partial w_{1}= (\partial E_{o1}/\partial out_{h1}+ \partial E_{o2}/\partial out_{h1}) * \partial out_{h1}/\partial in_{h1} * \partial in_{h1}/\partial w_{1}$
$\partial E_{total}/\partial w_{1}=(\partial E_{o1}/\partial out_{h1}+ \partial E_{o2}/\partial out_{h1}) * out_{h1}*(1-out_{h1})*i_{1}$     -(D)

Considering the 1st term in (C)
$\partial E_{total}/\partial out_{h1}= \partial E_{o1}/\partial out_{h1}+ \partial E_{o2}/\partial out_{h1}$

Now
$\partial E_{o1}/\partial out_{h1} = \partial E_{o1}/\partial out_{o1} *\partial out_{o1}/\partial in_{01} * \partial in_{o1}/\partial out_{h1}$
$\partial E_{o2}/\partial out_{h1} = \partial E_{o2}/\partial out_{o2} *\partial out_{o2}/\partial in_{02} * \partial in_{o2}/\partial out_{h1}$

which gives the following
$\partial E_{o1}/\partial out_{o1} *\partial out_{o1}/\partial in_{o1} * \partial in_{o1}/\partial out_{h1} =-(target_{o1}-out_{o1}) *out_{o1}(1-out_{o1})*w_{5}$ – (E)
$\partial E_{o2}/\partial out_{o2} *\partial out_{o2}/\partial in_{02} * \partial in_{o2}/\partial out_{h1} =-(target_{o2}-out_{o2}) *out_{o2}(1-out_{o2})*w_{6}$ – (F)

Combining (D), (E) & (F) we get
$\partial E_{total}/\partial w_{1} = -[(target_{o1}-out_{o1}) *out_{o1}(1-out_{o1})*w_{5} + (target_{o2}-out_{o2}) *out_{o2}(1-out_{o2})*w_{6}]*out_{h1}*(1-out_{h1})*i_{1}$

This can be represented as
$\partial E_{total}/\partial w_{1} = -\sum_{i}[(target_{oi}-out_{oi}) *out_{oi}(1-out_{oi})*w_{j}]*out_{h1}*(1-out_{h1})*i_{1}$

With this derivative a new value of $w_{1}$ is computed
$w_{1}^{+} = w_{1} - \alpha * \partial E_{total}/\partial w_{1}$

Hence there are 2 important results
At the output layer we have
a) $\partial E_{total}/\partial w_{j}=-(target_{oi} - out_{oi}) * out_{oi} *(1-out_{oi}) * out_{hi}$
At each hidden layer we compute
b) $\partial E_{total}/\partial w_{k} = -\sum_{i}[(target_{oi}-out_{oi}) *out_{oi}(1-out_{oi})*w_{j}]*out_{hk}*(1-out_{hk})*i_{k}$

Backpropagation, was very successful in the early years,  but the algorithm does have its problems for e.g the issue of the ‘vanishing’ and ‘exploding’ gradient. Yet it is a very key development in Neural Networks, and  the issues with the backprop gradients have been addressed through techniques such as the  momentum method and adaptive learning rate etc.

In this post. I derive the weights at the output layer and the hidden layer. As I already mentioned above, I intend to implement a vector version of the backpropagation algorithm in Octave, R and Python in the days to come.

Watch this space! I’ll be back

P.S. If you find any typos/errors, do let me know!

References
1. Neural Networks for Machine Learning by Prof Geoffrey Hinton
2. A Step by Step Backpropagation Example by Matt Mazur
3. The Backpropagation algorithm by R Rojas
4. Backpropagation Learning Artificial Neural Networks David S Touretzky
5. Artificial Intelligence, Prof Sudeshna Sarkar, NPTEL

To see all my posts go to ‘Index of Posts

# Neural Networks: On Perceptrons and Sigmoid Neurons

Neural Networks had their beginnings in 1943 when Warren McCulloch, a neurophysiologist, and a young mathematician, Walter Pitts, wrote a paper on how neurons might work.  Much later in 1958, Frank Rosenblatt, a neuro-biologist proposed the Perceptron. The Perceptron is a computer model or computerized machine which is devised to represent or simulate the ability of the brain to recognize and discriminate. In machine learning, the perceptron is an algorithm for supervised learning of binary classifiers

Initially it was believed that  Perceptrons were capable of many things including “the ability to walk, talk, see, write, reproduce itself and be conscious of its existence.”

However, a subsequent paper by Marvin Minky and Seymour Papert of MIT, titled “Perceptrons” proved that the Perceptron was truly limited in its functionality. Specifically they showed that the Perceptron was incapable of producing XOR functionality. The Perceptron is only capable of classification where the data points are linearly separable.

Checkout my book ‘Deep Learning from first principles: Second Edition – In vectorized Python, R and Octave’. My book starts with the implementation of a simple 2-layer Neural Network and works its way to a generic L-Layer Deep Learning Network, with all the bells and whistles. The derivations have been discussed in detail. The code has been extensively commented and included in its entirety in the Appendix sections. My book is available on Amazon as paperback ($18.99) and in kindle version($9.99/Rs449).

This post implements the simple learning algorithm of the ‘Linear Perceptron’ and the ‘Sigmoid Perceptron’.  The implementation has been done in Octave. This implementation is based on “Neural networks for Machine Learning” course by Prof Geoffrey Hinton at Coursera

Perceptron learning procedure
z = ∑wixi  + b
where wi is the ith weight and xi is the ith  feature

For every training case compute the activation output zi

• If the output classifies correctly, leave the weights alone
• If the output classifies a ‘0’ as a ‘1’, then subtract the the feature from the weight
• If the output classifies a ‘0’ as a ‘1’, then add the feature to the weight

This simple neural network is represented below

Sigmoid neuron learning procedure
zi = sigmoid(∑wixi  + b)
where sigmoid is
$sigmoid(z) = 1/1+e^{-z}$

Hence
$z_{i} = 1/1+e^{-(\sum w_{i}x_{i}+b)}$
For every training case compute the activation output zi

• If the output classifies correctly, leave the weights alone
• If the output incorrectly classifies a ‘0’ as a ‘1’ i.e. $z_{i} >sigmoid(0)$, then subtract the feature from the weight
• If the output incorrectly classifies a ‘1’ as ‘0’ i.e., i.e $z_{i} < sigmoid(0)$, then add the feature to the weight
• Iterate till errors <= 1

This is shown below

I have implemented the learning algorithm of the Perceptron and Sigmoid Neuron in Octave. The code is available at Github at Perceptron.

1. Perceptron execution

I performed the tests on 2 different datasets

Data 1

Data 2

2. Sigmoid Perceptron execution
Data 1 & Data 2

It can be seen that the Perceptron does work for simple linearly separable data. I will be implementing other more advanced Neural Networks in the months to come.

Watch this space!