Introduction
In this article, we will look at supervised learning algorithm called Multi-Layer Perceptron (MLP) and implementation of single hidden layer MLP.
Perceptron
A perceptron is a unit that computes a single output from multiple real-valued inputs by forming a linear combination according to its input weights and then possibly putting the output through some nonlinear function called the activation function.
Below is a figure illustrating the operation of perceptron:
[Figure taken from this link.]
The output of perceptron can be expressed as:
$f(x) = G( W^T x+b)$
- (x) is the input vector
- ((W,b)) are the parameters of perceptron
- (f) is the non linear function
Multi Layer Perceptron
The MLP network consists of input, output and hidden layers. Each hidden layer consists of numerous perceptrons which are called hidden units.
Below is a figure illustrating a feed forward neural network architecture for Multi Layer perceptron.
[Figure taken from]
A single-hidden layer MLP contains an array of perceptrons.
The output of hidden layer of MLP can be expressed as a function:
(f(x) = G( W^T x+b))
(f: R^D \rightarrow R^L),
where D is the size of input vector (x)
(L) is the size of the output vector
(G) is activation function.
In case the activation function G is a sigmoid function, then a single-layer MLP consisting of just the output layer is equivalent to a logistic classifier:
$\begin{align} f_{i}(x)=\frac{e^{W_{i}x+b_{i}}}{\sum_{j} e^{W_{j}x+b_{j}}} \end{align}$
Each unit of input layer corresponds to element of input vector.
Each output unit of logistic classifier generates a prediction probability that input vector belongs to a specified class.
Feed Forward Neural Network
Let us first consider the most classical case of a single hidden layer neural network.
The number of inputs to hidden layer is (d) and number of outputs of hidden layer are (m).
The hidden layer performs mapping of vector of dimensionality (d) to vector of dimensionality (m).
Each unit of hidden layer of a MLP can be parameterized by a weight matrix and bias vector (W,b) and an activation function (\mathcal{G}). The output of a hidden layer is activation function applied to linear combination of input and weight vector.
Dimensionality of weight matrix and bias vector are determined by desired number of output units.
If the number of inputs to hidden layer/dimensionality of input is (\mathcal{M}) and number of outputs is (\mathcal{N}), then dimensionality of weight vector in (\mathcal{NxM}) and that of bias vector is (\mathcal{N}x1).
We can consider that hidden layer consists of (\mathcal{N}) hidden units, each of which accepts a (\mathcal{M}) dimensional vector and produces a single output.
The output is the affine transformation of the input layer followed by the application of function $f(x)$, which is typically a non linear function like sigmoid of inverse tan hyperbolic function.
The vector valued function (h(x)) is the output of the hidden layer.
$ h(x) = f(W^T x + c ) $
The output layer of MLP is typically Logistic regression classifier, if probabilistic outputs are desired for classification purposes in which case the activation function is the softmax regression function.
Single Hidden Layer Multi Layer Perceptron's
Let,
- (h_{i-1}) denote the input vector to the i-th layer
- (h_{i}) denote the output vector of the i-th layer
- (h_{0})=x is vector that represents input layer
- (h_{n}=y) is output layer which produces the desired prediction output
- (f(x)) denote the activation function
Thus, we denote the output of each hidden layer as:
$h_{k}(x) = f(b_{k} + w_{k}^T h_{i-1}(x)) = f(a_{k}) $
Considering sigmoid activation function, gradient of function wrt arguments can be written as:
$\begin{align} \frac{\partial \mathbf{h}_{k}(x) }{\partial \mathbf{a}_{k}}= f(a_{k})(1- f(a_{k})) \end{align}$
The computation associated with each hidden unit $(i)$ of the layer can be denoted as:
$h_{k,i}(x) = f(b_{k,i} + W_{k,i}^T h_{i-1}(x)) = f(a_{k}(x))$
The output layer is a Logistic regression classifier. The output is a probabilistic output denoting the confident that input belongs to the predicted class. The cost function defined for the same is defined as negative log likelyhood over the training data:
$L = -log (p_{y}) $
The idea is to maximize (p_{y}= P( Y =y_{i} \| x )) as estimator of conditional probability of the class (y) given that input is (x). This is the cost function for training algorithm.
Back-Propagation Algorithm
The Back-Propagation Algorithm is recursive gradient algorithm used to optimize the parameters MLP wrt to defined loss function. Thus our aim is that each layer of MLP the hidden units are computed so that cost function is maximized.
Like in logistic regression, we compute the gradients of weights wrt to the cost function. The gradient of the cost function wrt all the weights in various hidden layers are computed. Standard gradient based optimization is performed to obtain the parameters that will minimize the likelyhood function.
The output layer determines the cost function. Since we are using Logistic regression as output layer. The cost function is the softmax function. Let L denote the cost function.
There is nothing different we do in back-propagation algorithm that any other optimization technique. The aim is to determine how the weights and biases change in the network.
$ \begin{align} \frac{\partial L}{\partial W_{k,i,j} } \text{ and } \frac{\partial L}{\partial b_{k,i,j} } \end{align}$.
output layer
$\begin{align} L = -log ( f(a_{k,i}) ) \end{align}$
$\begin{align} \frac{\partial L }{\partial \mathbf{a}_{k,i}} = \frac{\partial L }{\partial \mathbf{h}_{k,i}} \frac{\partial \mathbf{h}_{k,i} }{\partial \mathbf{a}_{k,i}} = -\frac{1}{h_{k,i}} * h_{k,i}*(1-h_{k,i}) = (h_{k,i}-1)\end{align} $
$ \begin{align} \frac{\partial L }{\partial \mathbf{a}_{k,i}} =\mathbf{h}_{k,j} - 1_{y=y_{i}} \end{align}$
The above expression can be considered as the error in output. When (y=y_{i}) the error is (1-p_{i}) and then (y \ne y_{i}) the error in prediction is (p_{i}).
hidden layer
$\begin{align}\frac{\partial L }{\partial \mathbf{a}_{k-1,j}} = \frac{\partial L }{\partial \mathbf{h}_{k-1,j}} \frac{\partial \mathbf{h}_{k-1,j} }{\partial \mathbf{a}_{k-1,j}} \end{align}$
Thus the idea is to start computing gradients from the bottom most layer. To compute the gradients of the cost function wrt parameters at the i-th layer, we need to know the gradients of cost function wrt parameters at (i+1)th layer.
We start with gradient computation at the logistic classifier level. They propagate backwards, updating the parameters at each layer.
Let us consider the case of other other hidden layers.
$\begin{align} \frac{\partial L }{\partial \mathbf{h}_{k-1,j}} = \sum_i \frac{\partial L }{\partial \mathbf{a}_{k,i}}\frac{\partial \mathbf{a}_{k,i} }{\partial \mathbf{h}_{k-1,j}} = \sum_i \frac{\partial L }{\partial \mathbf{a}_{k,i}} W_{k,i,j} \end{align} $
The implementation of the above equation is:
def linear_gradient(self,weights,error)
return numpy.dot(error,weights);
The gradients computation of parameters of hidden layers is as follows:
$\begin{align}\frac{\partial L }{\partial \mathbf{W}_{k-1,i,j}} = \frac{\partial L }{\partial \mathbf{a}_{k-1,j}} \frac{\partial \mathbf{a}_{k-1,j} }{\partial \mathbf{W}_{k-1,i,j}}=\frac{\partial L }{\partial \mathbf{a}_{k-1,j}} \mathbf{h}_{k-2,j} \end{align}$
$\begin{align}\frac{\partial L }{\partial \mathbf{b}_{k-1,i}} = \frac{\partial L }{\partial \mathbf{a}_{k-1,i}} \frac{\partial \mathbf{a}_{k-1,i} }{\partial \mathbf{b}_{k-1,i}}=\frac{\partial L }{\partial \mathbf{a}_{k-1,i}} \end{align}$
This is implemented as below, where the input:
- (x) represents (\begin{align} \frac{\partial \mathbf{h}_{k,j} }{\partial \mathbf{a}_{k,j}} \end{align}) -output gradient
- (y) represents (\begin{align} h_{k-2,j} \end{align}) -activation
- (w) represents (\begin{align} \frac{\partial L }{\partial \mathbf{a}_{k-1,i}}\end{align}) -error
def compute_error(self,x,w,y)
x=x*w;
res1=x.reshape(x.shape[0],1);
res=np.dot(res1,y.reshape(y.shape[0],1).T);
self.eta=0.0001
if self.Regularization==2:
res=res+self.eta*self.W;
if self.Regularization==1:
res=res+self.eta*np.sign(self.W);
res=np.hstack((res,res1));
return res.T;
def cost_gradients(self,weights,activation,error)
we=self.linear_gradient(weights,error)
ag=self.activation_gradient()
e=[ self.compute_error(a,we,b) for a,b in izip(ag,activation)]
gW=np.mean(e,axis=0).T
return gW;
Once we have the gradients and have computed the new parameters, the `function update` is called to update the new parameters in the model.
This function is called by the Optimizer module that performs SGD based optimizations, all the optimization parameters like learning rate are handled by the optimizer methods.
def update_parameters(self,params)
self.params=params;
param1=self.params.reshape(-1,self.nparam);
self.W=param1[:,0:self.nparam-1];
self.b=param1[:,self.nparam-1];
Implementation Details
The class HiddenLayer encapsulates all the methods for prediction, classification, training, gradient computation and error propagation that are required.
The important attributes of the HiddenLayer class are:
Attributes
-----------
`out` : array-like ,shape=[n_out]
The output of hidden layer
`params`:array-like ,shape=[n_out,n_in+1]
parameters of hidden layer
`W,b`:array-like,shape=[n_out,n_int],shape=[n_out,1]
parameters in the form of weight matrix and bias vector characterizing
the hidden layer
`activation`:function
the non linear activation function
.. note :
in the below functions to n_hidden denotes the number of output units of present hidden layer
n_out denotes the number of output units of next hidden layer
and n_in denotes the size of input vector to present hidden layer
def compute(self,input)
linout=numpy.dot(self.W,input.T)+np.reshape(self.b,(self.b.shape[0],1));
self.output=self.activation(linout).T;
return self.output;
A class MLP encapsulates all the methods for prediction, classification, training, forward and back propagation, saving and loading models, etc.
Below three important functions are displayed. The learn function is called at every optimizer loop.
This calls the forward and backward iteration methods and updates the parameters of each hidden layer.
The forward iteration simply computes the output of network and while propagate_backward function is responsible for passing suitable inputs and weights to each hidden layer so that it can execute the backward algorithm loop.
The forward iteration simply computes the output of network and while propagate_backward function is responsible for passing suitable inputs and weights to each hidden layer so that it can execute the backward algorithm loop:
def propagate_backward(self,error,weights,input)
input1=input;
for i in range(self.n_hidden_layers):
prev_error=np.inf;
best_grad=[];
for k in range(1)
hidden_layer=self.hiddenLayer[self.n_hidden_layers-i-1];
hidden_layer.compute(input1);
grad=hidden_layer.cost_gradients(weights,input1,error);
res=self.update(hidden_layer.params,grad.flatten(),0.13);
""" update the parameters """
hidden_layer.update_parameters(res);
weights=hidden_layer.W;
error=grad[:,hidden_layer.n_in];
self.hiddenLayer[self.n_hidden_layers-i-1]=hidden_layer;
input1=hidden_layer.output;
def propagate_forward(self,input)
self.predict(input)
def learn(self,update)
x,y=self.args;
self.update=update;
self.propagate_forward(x)
args1=(self.hidden_output,y);
self.logRegressionLayer.set_training_data(args1);
[params,grad]=self.logRegressionLayer.learn(update);
self.logRegressionLayer.update_params(params);
error=grad;
weights=self.logRegressionLayer.W;
if self.n_hidden_layers >0:
weights=self.logRegressionLayer.W;
self.propagate_backward(error,weights,x)
return [None,None];
Selecting the Parameters of the Model
As mentioned earlier, MLP consists of input, hidden and output layers. There is no fixed rule to determine the number of hidden units. The parameters are application specific and best parameters are often arrived at by empirical testing process. Less number of hidden units leads to increased generalization and training error while having a large number of training units leads to issues with training of large number of parameters and significantly large training time.
Issues with MLP
One of the issues observed in MLP training is the slow nature of learning. The below figure illustrates the nature of learning process when a small learning parameter or improper regularization constant is chosen. Various adaptive methods can be implemented which can improve the performance, but slow convergence and large learning times is an issue with Neural networks based learning algorithms.
Code
The important files related to MLP are:
- MLP.py
- LogisticRegression.py
- Optimizer.py
The latest version of the code can be found in Github repository at www.github.com/pi19404/pyVision.
The files used in the current article can be downloaded from the below link:
The dataset and model file can be found under the models and data repository:
- MLP.pyvision - model file
- mnist.pkl.gz - data file
Make suitable changes to the path in MLP.py file before running the code.
