is a veryvery important step it is women in spaceactually apre-processing step where we try to makesure you know we can uh move around thedata rearrange the data and we can makesure to refit and pack this data intoone particular range this is the entireprocess of normalization why we do thisis to make sure that you know if yourdata is in one particular range it meansthat your network has the ability tolearn very effectively and be veryefficient as well so uh you know whenyou're performing steps like backpropagation having a data in a very uhstructured range a specific range willmake sure that your network and learnthat much faster as well so these areone of two reasons why uh you know thispre-processing step that we take calledas data normalization is very importantnow with this we can come to the 16thquestion uh that states what is forwardpropagation so backward propagation isthe scenario where you know we work fromthe output layer to the input layer andhere in forward propagation it is theexact reverse we follow in the order ofthe neural network so uh you know we gofrom the input layers to the hiddenlayers and then uh the output layer soyou know when you're talking about acomplex neuro Network there'll be morethan one hidden layer so for everyhidden layer uh the output of theactivation function is actuallycalculated and uh you know this is doneuntil the next layer can actually beprocessed the reason they call it uhforward propagation as I just mentionedis to make sure we work from the inputand towards the output and this is theexact opposite of the backwardpropagation algorithm as well so to makesure that we use the output of theactivation function and to calculate itfor the next layers uh you know forwardpropagation is very very important nowwith this we can check out the 17thquestion uh again coming back to thevery main important question you knowyour interview could start with thisquestion itself uh you know what is backpropagationso back propagation you know instead oftalking about what the first thing youhave to always discuss is why backpropagation is even important right soit is to make sure that we keep the lossfunction or the cost function less andto see how the values changes wheneverwe you know work with a variety ofbiases a variety of Weights whenever wechange all of these uh even a slightchange can have a very big impact onyour model to understand how uh thesework we require the back propagationit's called back propagation of coursebecause it starts from the output layerWorks backwards uh towards the inputlayer Direction uh so how is backpropagation calculated well it is veryeasily calculated whenever you work uhyou know by understanding the gradientthat is calculated at every singlehidden layer itself so it's as easy andstraightforward as that now with this wecan check out the 18th question thatstates what are hyper parameters in deeplearning so when you talk abouthyperparameters hyper parameters eventhough the term is very complex they'renothing but variables that are actuallyused uh in deep learning uh to actuallydetermine the structure of a neuralnetwork uh so these variables are putinto place to make sure that you know wehave a easy way of understanding so manydifferent factors and parameters in theDeep learning model you know be it uhthe number of hidden layers that arepresent or the learning rate of themodel and much more so uh hyperparameters are very important variablesthat will give us a lot of insight intohow the model or the neural network isfunctioning as well so moving on to thenext question uh it states how can hyperparameters be trained in neural networksso training a hyperparameter is veryimportant and understanding how you canconvey this to the interviewer is alsovery important because these are longprocedures and your ability to conciselyanswer it is very important so you knowto train a hyper parameter will requirefour very important components it's thebat size the number of epoch and there'ssomething called as the momentum andthere's a metric called as the learningrate so what happens in batch size isthat we'll take the input we'll split itinto batches you know batch size can bevery depending on however uh it isrequired based on the efficiency and alot of other things these batches can becut into further sub batches as well uhif that is the requirement so instead ofpushing all the input into a concurrentpipeline we just make sure to dig it upinto batches and then vary it across nowwhen you talk about Epoch and Epoch isbasically a dention of how many timesthe training data is run through theneural networks where it can train sothe number of epoch is the number ofiterations at which uh you know themodel uh the neural Network trains aswell so it's an iterative process thenumber will depend on uh you know somany different factors doesn't mean thatif you have a large number of epoch atyour model uh will be very very accuratewith every single iteration it'll go upto a point and then of course your modelwill work in fact in the oppositedirection now when you talk aboutmomentum momentum is actually put intoplace where we have an idea to check outwhat happens in the next consecutivesteps you know whenever data is being uhmoved around in a neural network it isvital to understand what happens happensnext right so uh you know there's nopoint if your model is very effic veryinefficient where it's passing arounddata and not uh you know driving anymomentum from it so momentum is thisparticular track where we make sure uhwe understand what's happening next aswell now coming to the last componentit's a very important component it's thelearning rate learning rate is what weare all after when you work with deeplearning because this is used as a veryimportant parameter uh where we tryingto denote what is the time that'srequired to actually update all theparameters when your model learns andwhen your neural network is ready totake in certain inputs and give youmeaningful outputs as well so these youknow four form to be very very vitalvery important components in the worldof deep learning uh whenever we go

(Editor in charge: web3)