00:00Hello everyone, welcome back to the lecture series on foundations of deep learning concepts and
00:06applications. In our previous lecture, you learned about single layer perceptron, a hands-on session
00:13in which you have learned how to implement a single layer perceptron model in Google CoLab.
00:19Together with that, you have got familiarized with the basics of the Google CoLab environment.
00:25In today's session, let us discuss about multi-layer perceptrons which is the foundation
00:31of deep learning models and artificial neural network. So, I will go to the contents for today's
00:38discussion session. The contents go like this, introduction to multi-layer perceptron,
00:45comparison between single layer and multi-layer perceptron, feed forward neural network and then
00:53we touch upon shallow versus deep neural networks. So, we will start with what is a multi-layer
01:00perceptron. So, multi-layer perceptron is a supervised feed forward neural network
01:06which is composed of multiple layers and these layers include an input layer, a hidden layer
01:14layer and an output layer. And each of these layers consist of a set of perceptrons which are called
01:22as neurons. So, if we see this picture, we have an input layer here, we have a hidden layer here
01:33and an
01:34output layer here. The input layer consists of the neurons and the inputs come to these neurons
01:43x1, x2, x3, x4. These are the inputs to these neurons in the input layer. Then we will see that
01:50there are similar neurons in the hidden layer as well and all the neurons are interconnected.
01:59That means, there is a connection between the first neuron in the input layer to all the neurons
02:05in the hidden layer. If you can see, all are connected and similarly for second neuron, third neuron and
02:11fourth neuron in the input layer. And then, these hidden layer neurons are all connected to the output
02:18layer and from the output layer, we get the output of the model. Now, in your last session, you all
02:28have
02:28learnt about single layer perceptron and today we are going to learn about multi-layer perceptron, the first
02:36very basic multi-layer perceptron you have seen in my previous slide. Now, we will have a comparison
02:42between a single layer perceptron and a multi-layer perceptron. So, if you see, this is a single layer
02:51perceptron where these are input neurons and these are the output neurons. And the connections between the input
03:00and the output neurons, they are having some weights which are nothing but the importance that you give to
03:07each and every input neuron. So, in case of single layer perceptron which is this image, there is no hidden
03:18layer.
03:19The input neurons are directly connected to the output neurons. Whereas, in multi-layer perceptrons,
03:28the input layers are connected to the neurons in the hidden layers and they are connected to output
03:35neurons. Okay. So, what happens is there is a hidden layer of neurons between the input and the output.
03:44And here also, these connections or the edges are having some weights and these weights decide how
03:51much importance to be given to each of these neurons, to the input neurons. Of course, from the hidden layers
03:57to the output neurons when they are connected, there also we have the weights which decides how much
04:02importance will be given to the input that is coming from this hidden neuron to the output neuron.
04:10Now, in the second figure, you see there is only one single layer of hidden neurons. That means one
04:16single hidden layer is there. Multi-layer perceptrons can have only one hidden layer or it can have
04:26multiple hidden layers. So, my third picture is having multiple hidden layers and it is also an example
04:32of a multi-layer perceptron model. So, in single layer perceptron, we do not have any hidden layer.
04:38Multi-layer perceptrons can have one hidden layer or multiple hidden layers. Now, the next question
04:47comes, why do we need multi-layer perceptron? What is the requirement? Multi-layer perceptrons are used
04:55because they can handle complex data sets which are not linearly separable. Now, when we tell about
05:02complex data sets which are not linearly separable, what are those data sets? Where exactly we want to
05:11use these multi-layer perceptrons? They can be used in image recognition like handwriting,
05:18handwriting, handwritten digit recognition where your input is an image. It can be used in natural
05:25language processing like sentiment analysis. These data are not linearly separable data, they are much
05:31more complex data and it can also be used in speech recognition like voice to text transcription and
05:39similar applications like that. And these multi-layer perceptrons, they have the ability to learn complex
05:46patterns from a very large volume of data which is an essential tool in machine learning. So, when the
05:55data set is large and which is the case in lot of our real world problems, we require these multi
06:02-layer
06:02perceptrons because they can handle large volume of data. Also, they can handle complex data sets like image,
06:09text, speech, text, speech etc. And they can solve lot of complex problems.
06:17Now, we will quickly have a glance into the working of a perceptron. This you have done in your previous
06:24sessions in single layer perceptrons also, but I will just glance it through once more. So, we have
06:31this input features x1, x2 to xn and the perceptron it receives n features as input x1, x2 to xn
06:45and these
06:46features are connected with some weights which once again I repeat that these weights decide how much of
06:54importance we give to each and every input. And once they come
Comments