Deep learning as the name sometimes gives a perception to beginners that it some very difficult and hard to understand field and makes a perception that only PHD's student can understand and do it. If somebody tells you that your perception in wrong and says that any man who has done math's in his higher school can learn it and code it, then what would you do ?
Yes deep learning can done by any one who has done maths in higher classing and has some experience of coding can code deep learning easily. In this post we will provides you steps which here needed in deep learning. It contains or needed only three things data, architecture and loss function, For any deep learning you need to figure out this three things and you are done with it. Machine will learn by itself by using this three components.
As the image above shows three components and hierarchic structure for modelling deep learning. We will discuss each one them in some details and steps to connect to other components.
First component which you needed for deep learning model is data. Data can be of any for like Image, Audio, Video, Tabular, SQL. All data needed to be first divided into three sets which are Trains, Validation and Test sets and need to convert all of them into numerical values because deep learning architecture only understand number. If you have some strings in the data you need to convert to numbers by some mean. After converting to numbers you need to divide your full data into batches of fixed size. Let say you have 6400 items in your training set and you have chosen 64 items to be batches in one. So we have 6400/64=100 batches in your train data set. We divide into batches so that learning can be done faster.
Second component in deep learning is architecture which you need to figure. Basically architecture is made up of layers and each is connecting to one other layers. Each layers contains weights and bias which is learned through training process. All weights and bias numbers should be of appropriate type and amount because if you provide more weights and bias than your model will over fit and provide less than your model will under fit. If provide each batch to architecture which calculates its output and pass to loss function. This calculating and passing of data from architecture to loss function is called FORWARD PASS. Layers which we use in architecture are convolution layers, Recurrent neural network layers, Linear layers Activation function layers which are RELU, Sigmoid, softmax and many more. If you didn't get this layers don't worry we will cover all layers in future posts.
Third component in deep learning is loss function. This is the function which calculate numerical value between the values which is provided by your architecture and labels which you have because of this value a backward pass is done which is given optimizer which updates and rectifies weights and bias through that loss numbers which is provided by this function. This components improve our model to perform well. Good loss function you have your model will be as good . Different used loss function are MSE (Mean Squared Error), binary-cross entropy, Categorical-entropy, Log loss.
After back propagating loss function, Function value to improves over weight and basis. We repeat full process again and again for batches after batches and epoch's after epoch's to improve our models. These are only basic step needed for deep learning programming.
in this post we have explore and provide steps to program deep learning model. As its name suggest it is not that difficult to learn or to achieve.