Topic 21 | Introduction to Tensorflow

 AI Free Basic Course | Lecture 21 | Introduction to TensorFlow

The stage we so far have studied is called conventional machine learning. In conventional machine learning, we work on features. In other words, the algorithm of machine learning would not work directly in case the raw data, image or unstructured data is given as an input. Conventional Machine Learning algorithm means the algorithms like Linear regression algorithm, Logistic regression algorithm, decision tree algorithm, random forest algorithm and boosting algorithm. It is called conventional machine learning, as it can’t process directly from the raw data. It is a fundamental limitation of Conventional Machine Learning. We used sklearn, which is a well known library of conventional machine learning. We discussed necessary techniques of data sciences in it. Furthermore, we did exploratory data analysis of structured data, that how can we clean it, how can we transform it to enable the model processing it to produce better results.

We first studied regressors which predicted scores and then in order to evaluate the score we discussed root squared errors measures, mean squared errors measures. Likewise, we have seen that these used the structured data or features as input.

We then discussed different classifiers and their made their comparisons that which classifier is performing better than the other.

In our last lecture we discussed evaluation matrix like confusion matrix which is base through which other matrix like accuracy, precision and recall evolve. FM measure on which you have to make research.

We spent much time over EDA in which we visualize data, involve in cleaning and scaling data. We do a lot of preprocessing on data in order to make it suitable for the model to work on it for desired output.

After Conventional Machine Learning, we are going to discuss the other branch of Machine Learning that is called Deep Learning.

In deep learning, the job which have been performed by regressors and classifiers so would now be performed by Neural Networks. 

So the branch of Machine Learning in which we make use of Neural Networks in order to perform different tasks is called Deep Learning.

What is Neuron. Neuron is the unit of calculation in our brain. Our brain consist of billions of neurons which collaborate with each other in order to perform the task. The idea of Neural Network has evolved from Biology.

A natural question arises in our mind that if the neurons can do computations, then why these computations could not be done by computer. Why the technique used by neurons to perform computation should not be used in artificial intelligence to do the same. This idea started working on since 1960. So the Deep Learning is not the new thing.

Our brain always makes decision. If you have to decide that the picture is of bus or car. In order to make this decision, our brain needs input. We need the brain to give input in the shape of an image to make a decision.

From the voice of the person as input, we can not only recognize the person but can also recognize the tone of voice, whether it is angry, happy or high or low.

The techniques or powers of the neurons in our brain used to identify or perform multiple tasks when applied to computers take the shape of artificial intelligence. The function of neurons when manipulated or programmed in computer becomes artificial neuron and the network so formed is called neural network.

By making use of network of neurons in our brains we perform multiple tasks like watching, listening and speaking. The neural network in our brain is an advance form of neural network, but the neural network used in computers has not yet made such advancements to perform multiple tasks. It can only identify the voice of a person or image of a person at one time. These two tasks could not be performed by artificial neural network simultaneously. The advancement in the field is being made, which will be studied accordingly.

There are a lot of similarities in our brain and artificial neural network.

As discussed earlier, that our machine learning models fail to perform after a certain limit. For example, the model we studied were linear models. The data given was in binary form which was separated through a line as in fig A above. The line which separates the two types of data is called decision boundary. This line helps to identify about the location or area of point A and Point B. It means that when the data is in simple form, the linear models easily segregate the data.

When the pattern in the data is complex as given in the figure. The data given could not be segregated between point A and B through a straight line.

The data in the diagram has been classified in circular pattern. So it means we are shifting towards nonlinear technique to classify our complex pattern data. In complex data, we will not be able to identify the pattern of data and process it so that the model could use it to train itself.

It means we have to adopt such techniques through which we not only get rid of data preprocessing but make responsibility of model to identify the complex pattern in the data and decide whether there is need to draw a circle or oval in order to classify the data into different classes.   

The diagram given above is showing the big difference between the Machine Learning and Deep Learning.

In machine learning there is existence of a human which describe the image of car and converts it into structured data enabling model to understand it.

In order to ease the process and avoid the hectic job of converting the data of millions of cars into structured data, in deep learning we shifted the job of converting the data into machine learning language to the model itself.

In deep learning the network of neurons is accepting the image, processing it and giving out put on basis of it.

As it is not possible to shift the advance neurons of brain into computer, so the simple functions of maths were named as neurons. For example, you have to identify the larger number if two number of given to you. The Max Function called Relo Neuron can do this job and it was named as neuron. You would not believe it is the one of popular neurons used in AI. So the neurons are the simple functions of mathematics. So in neural network we adjust Relo Type functions into different layers which process the input and give us the output.