# What is backpropagate Reiteration in Neural Networks?

If we take I have 540 variables so I do a square root off 500 lie could be said 22.36 and I’m rounding it off to 22 so I’ll have first hidden lead with 22 Newman’s Then I am doing a square root off Hey Needles That gives me 4.69 and if I rounded off to five so I’ll have a second hidden layer with five new Ron’s Then I do a square root of five square root off five which gives me 2.23 So I round it off to know for example So then I’ll get 1/3 Hey didn’t lead with um doing your guns And then when I’m doing this with our I give hidden equal store when Tito comma five comma.

This means three layers first having 20 Don’t You’re on second having five neurons thought having Don’t You’re nuts Okay um till next is um the threshold So their shoulders nothing but they’re innovative off the error that will be calculated that each step So we can also see that from the help in our age See the special little specialize No medical value specifying the threshold for the partial derivative of the error function as a stopping criterion So at each step some era will be calculated as a difference off the actual value and the desired value And then your calculator derivative off it.

Say the error in the first step is so for example so you calculated a derivative Elphick then bye propagation with locker than a reiteration with Legrand this process keeps repeating so this threshold equals to 0.1 sees that keep repeating until you receive or derivative of the aerial function which is 0.1 Okay So in the first on the force situation may be on the air Live off derivative Off the error function will be safe for example one Then you backpropagate Reiteration occurs then the derivative of the edit function maybe comes down to 0.9.

Then you repeat the backpropagation occurs then Ah maybe the that give off the other function comes down to be 0.6 10 0 point does you don’t find one and they left Comes down to 0.1 You need to keep repeating the card, Okay Um the next step is So this is um one stopping great baby A They’re shoulders one stopping criteria Step Max tells you what is the number of it rations You want it It’s possible that um hey went with 10,000 integrations or 10,000 cycles The threshold doesn’t read 0.1 doesn’t mean that you’ll keep planning this algorithm infinite times Step Max is another stopping right area Which tells us what is the number of iterations that we can allow one.

Your network to have The step Mexico’s to 22 2000 tells us that we need to um we can run this and got it on a maximum of 2000 times Okay um Selenia output equals two falls is nothing but um telling whether or not you want the activation function to work So ah tau revisit activation function was nothing but allowing a neural network to fire a trigger Firing a trigger means allowing in your network to backfire So very bad Proper get So when When new specify Lee near output equals two falls That means that you’re allowing the activation function Okay So with the lectures different activation functions would also give.

But there’s not much difference practically speaking with ah the different activation functions And even if you don’t specify it explicitly with the commands you know you’re not missing out anything This specified that Lee near output equals two falls This is enough to tell us that you want the activation function to book Okay um the life sign nickels to fallen Life signed out The peak was too tender again Optional steps Um these are only to tell our that Say I am telling here that I want 2000 integrations Eso These two steps are only to tell our that I want to see the output after every tent it race If you step If you skip this step even then deal garden build on.

It will not show you what is happening with each step So when I’m telling our that I want the life sign Step Toby 10 So after each then it rations Like after Tenet rations after 20 after 30 after 40 So until 2000 I will show me what is happy I mean with the new network if you skip does it directly give you win What Okay then um head or not If the city is nothing but are telling you how do you calculate the error function, Okay Um so the editor is a real function is nothing but a difference off the um actual value and the desired value But how do you can collect it is to see a sum of squares that are and this is the best you can do with neural networks Because with your networks you want your error Toby positive so that you’re able to do a derivative of it So the best thing to use an SS e because some off squares Andrew So supplying all these bana meters when I run this command.

So I ran this command See there are don’t hidden new runs That were what I would have specified Ah with each situation So I told that I can do a maximum of 2000 steps but my unguarded um stopped with 1 25 steps Ah so when they were 10 steps when penetration that occurred My error was 2.869 after 23 years to 2.795 after 32 adios to 1.47 And finally after 1 25 it rations my era reduce to 0.354 one Um so the stopping right area which worked here was this threshold So a dead innovative off 0.35 fun would have been lesser than 0.1 n hinted stopped here Okay um toe want theater off 0.3541 was achieved This was achieved with 1 25 steps so I did not have to do 2000 steps I got the output with each tenant rations.

You see you can You could have done a life sign Step off one to see what Hap opened with meditation comes to be bailed Eso with each penetration You see the error was reducing and finally it came down to 0.351 on this occurred in 17 seconds Okay Now if I kind of tried toe lot the noodle network this is what I get So see there are four input variables Zeppelin with better lantern but ah I have just one um hidden layer That has it’ll neurons Ah these are the weight The ones that you see 1.83 on the top Sick Yeah So this 1.3 that you see 1.83 that you see here is to wait.

These are all the weight that you seeing Um then zero The 6.6 that you see here The minus 12 that you see here, Sorry Ah these are the biases So 1.830 point 76 the one that you see on the black lines are the weights that are being applied to the input variables The ones that you see in blue 6.6 and 1.2 These are the biases which are being applied So there’s just one hidden layer with till neurons Devices are being applied that both the layers the base of the hidden the rates are being applied to both the labs and finally and output since being thrown out Okay Ah One thing to note here is um that if I ran the same command again the way it will be different Okay.

If you’re on the same commander the same time with me even then this will be different So far what is happening is um I designed a score in is a neural network model which is being created It’s just showing you that these were the four input variables are these were the three output variables and something is happening on this new network is getting trained Okay um one additional step which can be done Shores Um not with the small later said that I’m using But if you have larger data sets on one another parameter that you can pass with the neural network commanders Um this step of the spiral meter called start rates.