What is the MNIST dataset?

Lesson Details:

Video Transcription: Let's look at what the this dataset actually is this video is brought to you by Appy Pie's Academy okay so now in this dataset is basically a big data set or a set of images which are handwritten digits okay and next to those handwritten digits you also pass in labels of those particular digits okay so over here in the first row we have zeros which are handwritten zeros and then whenever you are passing this to a machine learning model which we'll come back to shortly then you also tell the model that this is an image of a zero okay so whenever you are passing in any particular import you also give the output value it should have been generated okay so over here it is fairly easy for the human eye to also classify some basic digits over here as immediately just by looking at these particular images we can tell that this is a particular digit zero or three four or eight or something of that sort but for a computer this is very hard because whenever a computer looks at any particular image it only sees a huge line of pixels and then those pixels are arranged in a manner in which you can relate a particular image but you can see the image but making sense out of that particular image is very difficult for a computer directly and that is when we have machine learning or artificial intelligence which comes into play.

Okay so now let's spend some time and understand what is the difference between a classical programming approach and machine learning programming approach and why a machine learning programming approach is most efficient to solve a particular problem like the MS data set which is further used to check the accuracy or to check how strong or how accurate is your model when looking at a particular set of images so that is how we are going to also use the MS data set to create a model which can train on the MS data set on the Google our console and then we look at the loss and also see some results okay so over here first of all we have two red boxes over here which is first of all the machine learning model and then we have the classical programming model okay so now first of all whenever you want you have a particular set of input and output so in this particular case the images the handwritten images are the handwritten images of the digits are our inputs and the label of what these images actually correspond to is our output so in the case of a machine learning model we pass in the input and the output as you can see over here so we are passing in the training labels training images and also the labels and the tasks of this particular machine learning model is to generate the rules for us okay so it generates weights and biases which are further other rules which when put into a neural network can give you an exact output.

Let’s look at a simpler version okay so over here in India we get voting rights when we are at the age of 18 okay so if you want to build a machine learning model which tells if you are capable to vote or not then that is when you have to somewhere put in this particular rule which says that if age is greater than 18 then you can vote and if age is less than 18 then you can't vote okay but if we were to write a machine learning program then we would pass in a set of parameters okay so we would pass in the age and also of value yes or no if - here if to that particular person can vote so if we pass in 17 and say no and that is when the machine learning model generates some particular set of rules then when we pass in 19 and we say yes then there are more rules which is generated and then slowly by passing in hundreds and thousands of examples the machine learning model slowly generates a good set of rules which you can then use to check if you are eligible to vote or not so in this particular case we have only passed in the age and the availability or the allowance to what and the machine learning model has given us the particular set of rules so in this particular case the rules which are generated are a little bit more complicated than a simple if statement which compares her age to 18 so in this particular case the set of rules which are generated are weights and biases so as we have looked in the neural network section to define a particular neural network you need all of the weights and the bias of each and every neuron so once you have the weights and the biases you can define a well-structured neural network okay so in this particular case the crux of the upper part of the diagram is that there is a machine learning model which takes in the input and the output and then generates the set of rules for us so then we come up to the next part of the equation or the model which is a classical programming model and this is also required for a program to function properly so now in the case of a classical programming model we pass in the texting images okay so we pass in some data set which is not in the initial data set which was used for training okay so we pass some new samples into this particular classical programming model then the we also take in the rules which were generated from the machine learning model above which are the weights and the biases and we feed these two things which are the rules and the input to the classical programming model and then at the output stage you get a particular output label okay so initially when you are looking at the process of training the model you will only be cycling through the upper half of this particular cycle where you will only be passing the training images and the labels and then checking the lost value then you will again run the entire entire program and you will make sure that the lost value is reduced by implementing various optimizers like gradient descent or atom optimizer which we'll come back to shortly but for now for the process of training all you have to do is to change the value of the rules to change these particular rules so that the lost value or your error rate from the expected output which is provided over here is the minimum okay so this completes the training part and then comes the testing part when you pass in a particular image and then you also pass in the rules and check the output value okay so whenever you are doing any Google search or you are using your phone for face unlock you are only using this part of the cycle which is the testing part okay so that is pretty much it for the M Ness data set I also wanted to extend this particular lecture to also talk about the entire cycle of how a machine learning program actually works and what are the different processes involved and yes that is pretty much it about the M Ness data set and now from the next lecture onwards I will go behind the screen and we will start programming a particular M Ness data set classifier .