50% butwhat I can ai tools for developersdo is that I can onlygradually update my weights and I canonly gradually come to the best possibleoutput right so this process will keepon happening right this process willkeep on repeating if you see this thisLoop right this Loop of your X1 W1getting multiplied bias getting addedactivation function getting appliedright this will keep on happening thiswill give you an output the output willgive you the error the error will takethis back the weights will be updatedand then this process will keep onhappening right and this we are going tosee in some time now this process isknown as forward propagation right inforward propagation what we are doing isthat we are multiplying the input valuesby weight values we are taking themthrough the activation function and weare getting an output that is forwardpropagation right once when we get anoutput we calculate the error and we tryto read distribute this error back tothe initial sets of Weights this isknown as backward propagation right thisonce when we'll start with our uh properneural networks at that point of timethese two concepts will come intopicture but uh for now I'm justexplaining this that these are theforward propagation is nothing but yourweights getting multiplied by thefeatures and getting the linear and thenonlinear transformation backwardpropagation is the calculation of errorcoming from the assigned weights andthen red Distributing this error to goahead and update the weights in order toget a lesser errorright indu when we say input they willbe see it depends upon the features ifyou have 10 features then your inputwill be 10 different values right so ifI want to predict the price of a houseand if I have three features right solet's say area of the house the age ofthe house and the location then my inputeach and every time will be these threefeaturesright which activation function dependsright you will have to go ahead and youwill have to uh identify we will have totest around different activationfunctions right now activation functionis just like now let's say when you'replaying bedminton there is a differentset of uh like uh interaction that youyou would try to do right and let's saywhen you are driving a car right solet's say bedminton your hand movementis free you can move your hand fromlet's say right from the back of yourhead to the front right but when you'redriving your car your hand movementshould be limited maybe the movementshould be limited to only 10% right 0 to10% movement but when you are playingbadminton the movement can be 0 to 100%now this will depend upon the activationfunction if you keep an activationfunction which is making the handmovements in the range of 0 to 100% fora driving session obviously you aregoing to go ahead and crash right if youmake the activation function which givesan output of 0 to 10% for bedminton youwill never be able to play bedmintoncorrectly right but this is somethingthat you are able to identify as in whenyou pick up a new skill right so justlike that over here the activationfunction is something that we'll have tofind out which activation function willwork the best overhere backward propagation is a feedbacklooponly right yes naan so backwardpropagation is a feedback loop alongwith the feedback so backwardpropagation is the part wherein actuallywe we apply gradient descent now all ofthese things we'll be studying uh uhstep by step but in gradient descent wealso try to find out the feedback loopwill tell us what is the error rightgradient descent approach will tell usfrom where from what all weights thiserror is actually coming up right sodon't worry about that as of now withtime we are going to understand each ofthoseConcepts each of each set of input isfor one instance so one row right let'ssay you have three features for one rowthey will go inside the perceptron rightthen you will get a prediction right youwill get an error error will getadjusted then the second row will go thesecond row will go once again the samething will happen right all right nowlet's come to the right hand side uhflowchart of this right if you see thesame points that I have explained rightwe have an input right we have a set offeatures for those sets of features weinitialize the sets of Weights bias andthreshold values right threshold whatare thresholds and all we are going tocover cover that in due time don't worryright uh C calculate the sum and passthrough an activation function I hopethis is clear to everyone so this willbe the linear and the nonlineartransformation and through that weproduce the output when we produce theoutput we calculate an error right ifthe output is correct we stop right butobviously in the first iteration theoutput will not be correct so what willhappen we will calculate the error rightwe know what is expected of us rightwhat was the actual output we know whatthe prediction is coming out to be wewill calculate the error aftercalculating the error we will update theweights right now this piece ofmathematics that you are seeing thisdifferentiation and all right don'tworry about this this is gradientdescent I will be explaining gradientdescent in detail there will be onesession only on gradient descent uh thisupdation that you're seeing is nothingbut gradient descent at play right howit works all of these things we'll becovering right now all you need tounderstand is that there is there willbe a technique called graded descentwhich will help us in updating theweights right once when the weights areupdated once again you come back to thisparticular stage wherein you calculatethe sum pass through an activationfunction now because my weights havechanged obviously the linear the outputfrom the linear and the nonlineartransformation should also change rightso if the output is changing obviouslyyour error will also change right errorideally should go down right so thisprocess if you see this Loop will go onand on and on until you produce thecorrectoutput okay and this is how the trainingof a perceptron or the training of aneural network happens uh make backwardpropagation happen after each and everyrow is passed you can make backwardpropagation happen

(Editor in charge: web3)