Tutorial 7- Vanishing Gradient Problem

[@kumarpiyush2169]

HI Krish.. dL/dW'11= should be [dL/dO21. dO21/dO11. dO11/dW'11] + [dL/dO21. dO21/dO12. dO12/dW'11] as per the last chain rule illustration. Please confirm

[@rahuldey6369]

...but O12 is independent of W11,in that case won't the 2nd term be zero?

[@RETHICKPAVANSE]

wrong bruh

[@ayushprakash3890]

we don't have the second term

[@Ajamitjain]

Can anyone clarify this? I too have this question.

[@grahamfernando8775]

@@Ajamitjain dL/dW'11= should be [dL/dO21. dO21/dO11. dO11/dW'11]

[@mahabir05]

I like how you explain and end your class "never give up " It very encouraging

[@manishsharma2211]

Yes

[@Xnaarkhoo]

many years ago in the college I was enjoy watching videos from IIT - before the mooc area, India had and still have many good teachers ! It brings me joy to see that again. Seems Indians have a gene of pedagogy

[@Vinay1272]

I have been taking a well-known world-class course on AI and ML since the past 2 years and none of the lecturers have made me so interested in any topic as much as you have in this video. This is probably the first time I have sat through a 15-minute lecture without distracting myself. What I realise now is that I didn't lack motivation or interest, nor that I was lazy - I just did not have lecturers whose teaching inspired me enough to take interest in the topics, yours did. You have explained the vanishing gradient problem so very well and clear. It shows how strong your concepts are and how knowledgeable you are. Thank you for putting out your content here and sharing your knowledge with us. I am so glad I found your channel. Subscribed forever.

[@tosint]

I hardly comment on videos, but this is a gem. One of the best videos explaining vanishing gradients problems.

[@lekjov6170]

I just want to add this mathematically, the derivative of the sigmoid function can be defined as: *derSigmoid = x * (1-x)* As Krish Naik well said, we have our maximum when *x=0.5*, giving us back: *derSigmoid = 0.5 * (1-0.5) --------> derSigmoid = 0.25* That's the reason the derivative of the sigmoid function can't be higher than 0.25

[@ektamarwaha5941]

COOL

[@thepsych3]

cool

[@tvfamily6210]

should be: derSigmoid(x) = Sigmoid(x)[1-Sigmoid(x)], and we know it reaches maximum at x=0. Plugging in: Sigmoid(0)=1/(1+e^(-0))=1/2=0.5, thus derSigmoid(0)=0.5*[1-0.5]=0.25

[@benvelloor]

@@tvfamily6210 Thank you!

[@est9949]

I'm still confused. The weight w should be in here somewhere. This seems to be missing w.

[@ltoco4415]

Thank you sir for making this misleading concept crystal clear. Your knowledge is GOD level 🙌

[@PeyiOyelo]

Sir or As my Indian Friends say, "Sar", you are a very good teacher and thank you for explaining this topic. It makes a lot of sense. I can also see that you're very passionate however, the passion kind of makes you speed up the explanation a bit making it a bit hard to understand sometimes. I am also very guilty of this when I try to explain things that I love. Regardless, thank you very much for this and the playlist. I'm subscribed ✅

[@amc8437]

Consider reducing playback speed.

[@gultengorhan2306]

You are teaching better than many other people in this field.

[@bhavikdudhrejiya4478]

Very nice way to explain. Learned from this video- 1. Getting the error (Actual Output - Model Output)^2 2. Now We have to reduce an error i.e Backpropagation, We have to find a new weight or a new variable 3. Finding New Weight = Old weight x Changes in the weight 4. Change in the Weight = Learning rate x d(error / old weight) 5. After getting a new weight is as equals to old weight due to derivate of Sigmoid ranging between 0 to 0.25 so there is no update in a new weight 6. This is a vanishing gradient

[@ToqaGhozlan]

Many thanks for you , 9:25 the output is 0.004 even your explanation is the best THX

[@sapnilpatel1645]

so far best explanation about vanishing gradient.

[@rushikeshmore8890]

Kudos sir ,am working as data analyst read lots of blogs , watched videos but today i cleared the concept . Thanks for The all stuff

[@satyadeepbehera2841]

Appreciate your way of teaching which answers fundamental questions.. This "derivative of sigmoid ranging from 0 to 0.25" concept was nowhere mentioned.. thanks for clearing the basics...

[@mittalparikh6252]

Look for Mathematics for Deep Learning. It will help

[@classictremonti7997]

So happy I found this channel! I would have cried if I found it and it was given in Hindi (or any other language than English)!!!!!

[@deepthic6336]

I must say this, normally I am kinda person who prefers to study on own and crack it. Never used to listen to any of the lectures till date because I just don't understand and I dislike the way they explain without passion(not all though). But, you are a gem and I can see the passion in your lectures. You are the best Krish Naik. I appreciate it and thank you.

[@piyalikarmakar5979]

One of the best vedio on clarifying Vanishing Gradient problem..Thank you sir..

[@marijatosic217]

I am amazed by the level of energy you have! Thank you :)

[@vikrantchouhan9908]

Kudos to your genuine efforts. One needs sincere efforts to ensure that the viewers are able to understand things clearly and those efforts are visible in your videos. Kudos!!! :)

[@al3bda]

oh my god you are a good teacher i really fall in love how you explain and simplify things

[@koraymelihyatagan8111]

Thank you very much, I was wandering around the internet to find such an explanatory video.

[@himanshubhusanrath2492]

One of the best explanations of vanishing gradient problem. Thank you so much @KrishNaik

[@MrSmarthunky]

Krish.. You are earning a lot of Good Karmas by posting such excellent videos. Good work!

[@skiran5129]

I'm lucky to see this wonderful class.. Tq..

[@yousufborno3875]

You should get Oscar for your teaching skills.

[@sumeetseth22]

Love your videos, I have watched and taken many courses but no one is as good as you

[@mittalparikh6252]

Overall got the idea, that you are trying to convey. Great work

[@aidenaslam5639]

Great stuff! Finally understand this. Also loved it when you dropped the board eraser

[@manujakothiyal3745]

Thank you so much. The amount of effort you put is commendable.

[@venkatshan4050]

Marana mass explanation🔥🔥. Simple and very clearly said.

[@YashSharma-es3lr]

very simple and nice explanation . I understand it in first time only

[@benvelloor]

Very well explained. I can't thank you enough for clearing all my doubts!

[@MauiRivera]

I like the way you explain things, making them easy to understand.

[@meanuj1]

Nice presentation..so much helpful...

[@elielberra2867]

Thank you for all the effort you put into your explanations, they are very clear!

[@classictremonti7997]

Krish...you rock brother!! Keep up the amazing work!

[@MsRAJDIP]

Tommorow I have interview, clearing all my doubts from all your videos 😊

[@adityashewale7983]

hats off to you sir,Your explanation is top level, THnak you so much for guiding us...

[@DEVRAJ-np2og]

do u completed his full playlist?

[@maheshsonawane8737]

Very nice now i understand why weights doesn't update in RNN. The main point is derivative of sigmoid is between 0 and 0.25. Vanishing gradient is associated with only sigmoid function. 👋👋👋👋👋👋👋👋👋👋👋👋

[@swapwill]

The way you explain is just awesome

[@vishaljhaveri6176]

Thank you, Krish SIr. Nice explanation.

[@prerakchoksi2379]

I am doing deep learning specialization, feeling that this is much better than that

[@nabeelhasan6593]

Very nice video sir , you explained very well the inner intricacies of this problem

[@b0nnibell_]

you sir made neural network so much fun!

[@sekharpink]

Derivative of loss with respect to w11 dash you specified incorrectly, u missed derivative of loss with respect to o21 in the equation. Please correct me if iam wrong.

[@sekharpink]

Please reply

[@ramleo1461]

Evn I hv this doubt

[@krishnaik06]

Apologies for the delay...I just checked the video and yes I have missed that part.

[@ramleo1461]

@@krishnaik06Hey!, U dnt hv to apologise, on the contrary u r dng us a favour by uploading these useful videos, I was a bit confused and wanted to clear my doubt that all, thank you for the videos... Keep up the good work!!

[@rajatchakraborty2058]

@@krishnaik06 I think you have also missed the w12 part in the derivative. Please correct me if I am wrong

[@yoyomemory6825]

Very clear explanation, thanks for the upload.. :)

[@aaryankangte6734]

Sir thank u for teaching us all the concepts from basics but just one request is that if there is a mistake in ur videos then pls rectify it as it confuses a lot of people who watch these videos as not everyone sees the comment section and they just blindly belive what u say. Therefore pls look into this.

[@benoitmialet9842]

Thank you so much, great quality content.

[@it029-shreyagandhi5]

Great teaching skills !!!

[@nola8028]

You just earned a +1 subscriber ^_^ Thank you very much for the clear and educative video

[@hiteshyerekar2204]

Nice video Krish.Please make practicle based video on gradient decent,CNN,RNN.

[@skviknesh]

I understood it. Thanks for the great tutorial! My query is: weight vanishes when respect to more layers. When new weight ~= old weight result becomes useless. what would the O/P of that model look like (or) will we even achieve global minima??

[@అరుణాచలశివ3003]

you are legend nayak sir

[@faribataghinezhad]

Thank you sir for your amazing video. that was great for me.

[@sunnysavita9071]

your videos are very helpful ,good job and good work keep it up...

[@shmoqe]

Great explanation, Thank you!

[@RAZONEbe_sep_aiii_0819]

There is a very big mistake at 4:14 sir, you didn't applied the chain rule correctly, check the equation.

[@jagritiprakash4336]

I have the same doubt

[@naresh8198]

crystal clear explanation !

[@susmitvengurlekar]

Understood completely! If weights hardly change, no point in training and training. But I have got a question, where can I use this knowledge and understanding I just acquired ?

[@tonnysaha7676]

Thank you thank you thank you sir infinite times🙏.

[@manikosuru5712]

As usual extremely good outstanding... And a small request can expect this DP in coding(python) in future??

[@krishnaik06]

Yes definitely

[@hokapokas]

Good job bro as usual... Keep up the good work.. I had a request of making a video on implementing back propagation. Please make a video for it.

[@krishnaik06]

Already the video has been made.please have a look on my deep learning playlist

[@hokapokas]

@@krishnaik06 I have seen that video but it's not implemented in python.. If you have a notebook you can refer me to pls

[@krishnaik06]

With respect to implementation with python please wait till I upload some more videos

[@muhammadarslankahloon7519]

Hello sir, why the chain rule explained in this video is different from the very last chain rule video. kindly clearly me and thanks for such an amazing series on deep learning.

[@daniele5540]

Great tutorial man! Thank you!

[@arunmeghani1667]

great video and great explanation

[@krishj8011]

Very nice series... 👍

[@spicytuna08]

you teach better than ivy league professors. what a waste of money spending $$$ on college.

[@BalaguruGupta]

Thanks a lot sir for the wonderful explanation :)

[@neelanshuchoudhary536]

very nice explanation,,great :)

[@salimtheone]

very well explained 100/100

[@Kabir_Narayan_Jha]

This video is amazing and you are amazing teacher thanks for sharing such amazing information Btw where are you from banglore?

[@magicalflute]

Very well explained. Vanishing gradient problem as per my understanding is that, it is not able to perform the optimizer job (to reduce the loss) as old weight and new weights will be almost equal. Please correct me, if i am wrong. Thanks!!

[@nirmalroy1738]

super video...extremely well explained.

[@abdulqadar9580]

Great efforts Sir

[@GunjanGrunge]

that was very well explained

[@nikunjlahoti9704]

Great Lecture

[@winviki123]

Could you please explain why bias is needed in neural networks along with weights?

[@Rising._.Thunder]

it is because when you want to control or fix the output of a given neuron within a certain range, for example, if the neuron is always giving inputs between 9 and 10, you can put a bias =-9 so as to make the neuron output between 0 and 1

[@nazgulzholmagambetova1198]

great video! thank you so much!

[@melikad2768]

Thank youuuu, its really great:)

[@dhananjayrawat317]

best explanation. Thanks man

[@gaurawbhalekar2006]

excellent explanation sir

[@khiderbillal9961]

thanks sir you really hepled me

[@gouthamkarakavalasa4267]

Gradient Descent will be applied on Cost function right ?-1/m Σ (Y*log(y_pred) + (1-y)* log(1-y_pred))... in this case if they had applied on the activation function, how the algo will come to global minima.

[@aishwaryaharidas2100]

Should we again add bias to the product of the output from the hidden layer O11, O12 and weights W4, W5?

[@naughtyrana4591]

Guruvar ko pranam🙏

[@narsingh2801]

You are just amazing. Thnx

[@gowthamprabhu122]

Can someone please explain why the derivative of each parent layer reduces ? i.e why does layer two have lower derivative of O/P with respect to its I/P?

[@karth12399]

Sir you are saying derivative of sigmoid is 0 to 0.25. I understand it. But how that will imply derivative of O21 /derivative of 011 should be less than 0.25. Could you please help me understand that assumption

[@rish_hyun]

He agreed that he did it wrong subconsciously I found his comment somewhere in this chat

[@jsridhar72]

The output of every neuron in a layer is the Sigmoid of weighted sum of input. Since sigmoid is applied as the activation function in every neuron(here O21 is output after applying sigmoid function), the derivative should be between 0 and 0.25.

[@sunnysavita9071]

very good explanation.

[@ambreenfatimah194]

Helped a lot....thanks

[@sowmyakavali2670]

Hi krish everyone says that Wnew = Wold - n * dL/dWold theoritically we know that dL/dWold means slope where as in practical scenario L is a single scalar value Wold is also a single scalar value Then how dL/dWold is calculating ??? And also coming to the activation function , you are explaining it theoritically , can you explain it by taking practical values ? , and don't tell it by taking predefined function or module , bcz we know how to find a module and import it and how to use it , but we don't know practical

[@LazingOnSunday]

This video is really goooooddd! Can anyone help me understand why the derivate value decreases as we go backward @9:03? I am new to DL..!!

[@vaseekaranchittibabu2571]

yes i have the same doubt. Can someone explain this?

[@shahidabbas9448]

Sir i'm really confusing about the actual y value please can you tell about that. i thought it would be our input value but here input value is so many with one predicted output

[@gautam1940]

This is an interesting fact to know. Makes me curious to see how ReLU overcame this problem

[@sandipansarkar9211]

Thanks krish .Video was superb but I am having apprehension I might get lost somewhere .Please provide some reading reference regrading this topic considering as a beginner.Cheers

[@ayushprakash3890]

is this equation correct ?? (this equation is used in the starting of the video) dL / dw11 = dO21 / dO11 * dO11 / dw11 should it be : dL / dw11 = dL / dO21 * dO21 / dO11 * dO11 / dw11

[@amitdebnath2207]

Hats Off Brother

[@Joe-tk8cx]

Great video, one question, when you calculate the new weights using the old weight - learning rate x derivative of loss with respect to weight, the derivative of loss wrt weight is that the sigmoid function ?

[@ashwinshetgaonkar6329]

output O21 also depends upon O12,so its derivative should also be considered

[@FlyingTurtleLP]

What I didnt get: What can the values of the derivative of the sigmoid function be? I don't think you mentioned it in the video.