Deep Learning 27: (1) Generative Adversarial Network (GAN): Introduction and Back-Propagation
Вставка
- Опубліковано 23 лют 2019
- In this lecture introduction to generative adversarial networks (GANs) is carried out in detail. The primary focus of this lecture is on working and back-propagation process.
#adversarial#generative#deeplearning
WOW!!! so far the first lecture is way better than any online resource I've seen!!! thanks for sharing!!!
The description of generator backprop at 16:55 was very helpful. Thanks.
Wow, Sir! Thank you for this wonderful video.
The explanation is very clear and simple. Thank you so much for this precious content, Sir! Please keep making more videos.
Most underrated channel ever...Your content is love man! Keep creating!!
Thank you so much for this amazing explanation.
Thaaaank you!! really nice illustration of how GANs work !!
Thank you so much for the video. Great and simple explanation.
Great and most precise explanation of deep learning I have seen. Can we have the links of these notebooks?
Amazing explanation!
Thank you sir for such a wonderful playlist . :)
Very good explanation of GAN.
Great explaination, I found it much useful
Sir! you are doing a great work... Thank you!
Thank you so much. Your Gan description was very helpful.
Amazing explanation, thank you
Excellent video 👍
Thank you for your explanation ... so clear.
Sir your way of explaining this complex topics are very much understandable lots of love from India🙏🏻
Really great Sir
Sir,your explanation is crystal clear
thank you sir ,these videos will help me in research process.
thanks alot for making these. really mean it!
amazingly explained.....understood the notations so so easily. its easy to read a research paper related to GAN comparitively better.
Superb helps a lot
THANK YOU!!!!!!!!! I FINALLY GET IT!
very clear Explanation..Thank you sir
Thank u soooo much!!! I was eagerly waiting for this for my research. You have seriously come in to my life as an angel at this time as i have been struggling a lot with my research topic and i dont wana get fail😪.
Hii, need one more help please.make a video of any working application of GAN method
thak you simply the best
Nice explanation sir. Pls share some projects on GAN
good and simple
Thank you sir 🙏.
Thankyou sir for the great content and explanation!!! It would be very helpful if you can share the notebook link with us. It would be helpful for revision
A W E S O M E explanation ...
Great videos . Please upload videos for CNN
Can you please make a tutorial on reinforcement learning and SeqGAN? It would really be very helpful
SIr, very nice explaination. Just one dubt:- Are you talking about some spatial distribution in the case of image or Distribution means simply the one which we study in statistics? Pl. reply. Thanks!
please make videos on deep dream and neural style transfer.
Nice video sir, however, at 16:40, I get that the discriminator will do the classification after the data is fed, but, will the model be trained on the combined data or would we do the prediction straight away?
In my research i need to explain math behind Gan method and implementation of Gan method and any working application of Gan method.i trust that ill get the needed knowledge from your videos...thanks a ton!!
yes you will for sure....
@@AhladKumar thank u so much!!🙏🙏🙏
Thank you prof. Could you add implementations in pytorch?
You are the best Kumar. I am new comer in GAN world, so I was looking for tutorials to help start the basics. This tutorial gave a global overview of Generative Adversarial Network (GAN): Introduction and Back-Propagation. Can I apply this for insect pest identification??????
gem video proff
A distribution means a cluster of photographs. G is learning algorithm generated. D is pre photographs learnt truths taught. Filtered D".
Which data augmentation method is best for time series data
Thank you for the wonderful video. I was a bit of confused, how the generator generates the image ? suppose in text to image synthesis we have a text which is encoded and joint with noise vector and they say the joint vector is passed to generator and image is generated . But this is being a black box to me ? What's the mechanism to convert the vectored sentence to image . Does the encoder does all by it'self ?
the generator will do it.....if you see closely its a neural network that takes noise vector of size let us say 10x1 and converts it to output vector of size same as that of input. It learns to create this kind mapping through back propagation....may be it will become more clear to you once you see other parts of my lecture series....happy learning !
Is there any ppt or document than we can follow for the video tutorials?
Hello sir,
Nice explanation sir
Can u plZ tell how we can use gan for data augmentation and deep learning alexnet /resnet-50/ vgg 16 for classification
sir any video which explains latent space easily?
Great video!
Couple of questions:
1: What are we achieving by making Generator generate D' close to D or by creating fake images if we can classify images correctly just using Discriminator.
2: How we get value 0.5 and not 0 or 1 after training both G and D. After training what would I expect as y for true image and for fake image.
thanks!
1.its not the generator actually ,its the distribution of the output of the generator pg we are making close to the distribution of the pdata of the original data . bcz the probability distribution of the fake images vary bcz they are not same in the high dimension , bcz the original images have a fixed distribution.
2. In order to maximize the objective of the generator i.e maximizing the js diveregence between the original data and the generated distribution data we have to find the local maxima of the function , so the local maxima is achieved only when the disciminator is 1/2 , this is achieved only when the differentiable function generated by the generator with parameters theta(g) are very close the parameters and distribution of the original data base .
can anybody explain me what is noise input to generator and why we use
it
sir wht do u mean by l2 nor , l1nor etc
can we have subtitles?
Hello sir can you please Guide for PHD on GAN
Is the latent space at 9:20 same as random distribution? Please, it would make things more clear for me.
yes
@@AhladKumar Can a short video or notes/synopsis or pdf doc on specifically be made on genrator? Discussion about architecture and loss function.Even if you put link of a blog or site which tells in laymans language will also do.
Also, big Thanks because of you I am on atleast level 5 wrt unsupervised models.
awesome
Can you share these slides
as close as possible to......D
can you please share the notebook. i mean these notes
Post the link for notes please
Can i apply this model for Time Series Data or stock price prediction
For stock price prediction you can use a recurrent network like LSTM and train it as a regressive model. you will need to feed it lots of different data to get any accuracy though
Don't we train Generator and Discriminator back and forth ? As you explained we simply first train D as a binary classifier, then G comes to picture and G(z) is fed to D and if D label it as fake then we backprop and update the weight of G only , Is it true? or there is the simultaneous training of both G and D is done ? I am bit confused.
Simultaneous training of G along with Discrim is not required since we already trained the Discrim to it's full accuracy so we keep it fixed..This is what i think in my opinion..if you found the correct answer please share it here.
@@MultiVedanth no, you don't train the discriminator to the fullest because it will be a huge issue for the generator. if it happens the discriminator will always be able to classify whether the generated images are true or not and the generator will not get any useful information from the discriminator to adjust its weights and biases. A strong discriminator and a weak generator will never work.
for my knowledge, the generator is trained more often than the discriminator. for example, generator weights are adjusted for 2 iterations while the discriminator adjusts for 1. this is something I'm not 100% sure so if there's someone who knows, please let us know.
can we get the ppts of this
Hi Sir, Do we have real world applications for GAN?
Yes there are many. Generating synthetic data when data is not freely available due to privacy concerns. Like in healthcare or fraud detection studies, people generate synthetic data and can then share synthetic data with public without any privacy concerns.
sir can u share notes of this topic
sir , where the z data points came? is this the training data point?
z are the samples coming from a known distribution like gaussian
@@AhladKumar Sir is it Gaussian or Random distribution..Love your videos sir..great _/|\_
z is the input distribution , it is given to the differential distribution function or the generator to generator Pg output , which is different . it is the prior probabilty that wil change according the theta(g)
I am writing to express my gratitude for the amazing content you have been producing on neural networks. Your videos have been incredibly helpful in my end of studies project for my masters degree. I have been working on a GANs project, which will be a comprehensive review about GANs, and I am currently focused on the implementation part where I hope to achieve better results with certain GANs variations.
I wanted to reach out to you and ask if you would be willing to offer some guidance or assistance with my project. Your expertise and knowledge in this area would be invaluable to me, and I would greatly appreciate any help or advice you can provide. If you are interested in assisting me with my project, I would be more than grateful to receive your help.
Thank you for taking the time to read my message, and I hope to hear from you soon.
I want this pdf if you can
why u called Z to g
Everything was very well explained! You deserve more views! - Do you hear me youtube-algorithm?!?!
подчерк красивый
Baigan
Indian accent without sub is impossible to be understood by another country!