all this time I was thinking if Probabilistic ML techniques have been lost. But this lecture has put a smile on my face. A perfect blend of adding uncertainty by way of defining Gaussian priors, overcoming the back propagation hurdle.. this is simply SUPERB. and kudos to you Ms. Ava.
Hands down the best lecturers I've seen after Andrew Ng on Deep learning. Ms Soleimany especially, is quite comprehensive in this video. Thanks a log Ava, Alexander and MIT team for putting this up.
You know what's interesting is that the GANs seem to have issues with symmetry. So the GAN generated faces they show alot, the easiest way to tell them from real images is by looking at the ears, and the teeth. For instance both B and C have one earring, and one earlobe larger than the other. GANs do things like lighting so well, but I guess that tends to be more continuous where as the ear symmetry is further spread apart, or understanding that each tooth doesn't have a general tooth look, but a specific look and it's important that they're all there.
This course is awesome and thank you so much for sharing it with us! I have one question though - why is it such a common practice to include equations without labeling all the variables? For example at 19:06 there is no mention (or maybe I missed it?) of what the "D()" function is, or what the || symbol represents.
D denotes the distance between the distribution that the encoder learns (i.e. p_phi(z|x) ) and the prior that we chose (i.e. p(z) ). I'd recommend you to see the lecture on it from last year by Amini himself : th-cam.com/video/yFBFl1cLYx8/w-d-xo.html
Thanks for sharing this great material. I come from a maths background and only recently got to dive a bit into ML and deep learning. One question that comes to my mind though - and this is more of a philosophical question than a practical I would guess - is why isn't there more focus on the QUALITY of the parameters learned by SGD. Mathematically speaking, we are only guaranteed a global optimum in the case of a convex loss function, so how do we evaluate the quality of learned parameters from local optima in the case of non-convex loss functions. Is there any research/mathematical research done that guarantees certain properties on these parameters, or their variability, or any other measure of stability for example? thanks.
I just have watched previous lectures on this topic, but I can't wait to see this lecture!!! Thank you a lot for such amazing content!!!!! Mind blowing! !!
Great video. I think the explanation of the difference between an autoencoder and variational autoencoder might be a little clearer with example output though.
For example I have some official covid data and some fake tweet and fb data of covid how can I model this to identify the policies have to taken by govt.
i wish i was smart enough to find a way to unify art & perception with deep neural networks. or a system that could help me interoperate brain activity. but i make no sense beyond my own mental limitations
In style transfer example (Cycle GANs) 39:00 ...where Alexander's speech style is transformed into Obama's style ...Why does Obama's lip sync match with spoken utterance ? When it's just speech to speech style transfer
Yes it is, they are a learnt model that does lossy compression (the encoder part, from the input to the smallest layer) and then decompression (the decoder part, from the layer just after the smallest one to the output).
@Alexander @Ava, you mention the Generator sees the real data, but we don't pass the Generator the real data anywhere, I thought only the Discriminator sees the real data, tad bit unclear on that, it would be great if you could clarify this.
So far, 23:57 in. VAEs and GANs are so sexy! They're amazingly attractive from a learning POV (point of view). Thanks to Ava for explaining this to us. I actually understand it.
Can anyone recommend me best online course for deep learning in python, i mean course should have detailed explanation in deep learning,specially in CNNs
you should check coursera deep learning by andrew yang, after that there is also tensorflow in practice, but NOTE: - you should already by familiar with python - you should know the deffrence between deep N, and machine learning so that you'll know what to learn
It's very kind of you to share the course for the researchers all over the world!
all this time I was thinking if Probabilistic ML techniques have been lost. But this lecture has put a smile on my face. A perfect blend of adding uncertainty by way of defining Gaussian priors, overcoming the back propagation hurdle.. this is simply SUPERB. and kudos to you Ms. Ava.
It's quite impressive of how such a complex topic is explained very accurately....Thank you for the lecture !
Thank you very much Ava, excellent presentation. God bless you.
Hands down the best lecturers I've seen after Andrew Ng on Deep learning. Ms Soleimany especially, is quite comprehensive in this video. Thanks a log Ava, Alexander and MIT team for putting this up.
This is the best explanation of a VAE I've ever seen/heard, THANK YOU for sharing!
Wow! While explaining regularization term how wonderfully you explained the use of bayesian statistics !
Thanks a lot for such a great explanation :)
Excellent lecture, thanks. Been trying to understand GAN implementation and now I do.
Thanks for illuminating these areas.
سپاس از شما و باعث خوشحالی که در اولین نتیجه ی جستجوی گوگل اومد-پایدار و پیروز باشین
You know what's interesting is that the GANs seem to have issues with symmetry. So the GAN generated faces they show alot, the easiest way to tell them from real images is by looking at the ears, and the teeth. For instance both B and C have one earring, and one earlobe larger than the other. GANs do things like lighting so well, but I guess that tends to be more continuous where as the ear symmetry is further spread apart, or understanding that each tooth doesn't have a general tooth look, but a specific look and it's important that they're all there.
What a wonderful series of lectures! Thoroughly enjoying. Many thanks to the instructors and MIT.
I enjoyed every bit of all the lectures.
Best deep learning refresher class I've come across
Refresher?!
Finally....I stuck in reparameterization for a long time.......
Thank you Ava for a great explanation of the subject! Really intuitive!
This course is awesome and thank you so much for sharing it with us! I have one question though - why is it such a common practice to include equations without labeling all the variables?
For example at 19:06 there is no mention (or maybe I missed it?) of what the "D()" function is, or what the || symbol represents.
D denotes the distance between the distribution that the encoder learns (i.e. p_phi(z|x) ) and the prior that we chose (i.e. p(z) ).
I'd recommend you to see the lecture on it from last year by Amini himself : th-cam.com/video/yFBFl1cLYx8/w-d-xo.html
|| means ‘or’ in most programming languages
Thanks for providing great introductory lectures to DL!
Thanks for sharing this great material. I come from a maths background and only recently got to dive a bit into ML and deep learning. One question that comes to my mind though - and this is more of a philosophical question than a practical I would guess - is why isn't there more focus on the QUALITY of the parameters learned by SGD. Mathematically speaking, we are only guaranteed a global optimum in the case of a convex loss function, so how do we evaluate the quality of learned parameters from local optima in the case of non-convex loss functions. Is there any research/mathematical research done that guarantees certain properties on these parameters, or their variability, or any other measure of stability for example? thanks.
I just have watched previous lectures on this topic, but I can't wait to see this lecture!!! Thank you a lot for such amazing content!!!!! Mind blowing! !!
Great video. I think the explanation of the difference between an autoencoder and variational autoencoder might be a little clearer with example output though.
I have a plan now that I generate some fake with GaN and finally try to classify using SVM and hopefully margin will tell me goodness
For example I have some official covid data and some fake tweet and fb data of covid how can I model this to identify the policies have to taken by govt.
Thank you mam good Explaination of GAN & Autoencoder
Thanks for the updated lectures. I haven't watched any MIT deep learning lectures as of yet, but I am looking forward to it.
i wish i was smart enough to find a way to unify art & perception with deep neural networks. or a system that could help me interoperate brain activity. but i make no sense beyond my own mental limitations
In style transfer example (Cycle GANs) 39:00 ...where Alexander's speech style is transformed into Obama's style ...Why does Obama's lip sync match with spoken utterance ? When it's just speech to speech style transfer
I have the same question. I think they might have used another model for it but its just a guess. Who knows if they actually had Obama say that xD
Thank you for the concrete lessons. Any chance I can access the MIT lab contents including the lecture?
awesome and explained very well..
Would it be correct to say that autoencoders are a form of lossy compression?
Yes it is, they are a learnt model that does lossy compression (the encoder part, from the input to the smallest layer) and then decompression (the decoder part, from the layer just after the smallest one to the output).
This is awesome!
I have a doubt Mr Alexander. Is Gan and reinforcement learning be used side by side.
@Alexander @Ava, you mention the Generator sees the real data, but we don't pass the Generator the real data anywhere, I thought only the Discriminator sees the real data, tad bit unclear on that, it would be great if you could clarify this.
Deep Generative Modeling, Is this lecture 4 ?
what if we want to learn the latent space in GANS how can we embed VAE to that
So far, 23:57 in. VAEs and GANs are so sexy!
They're amazingly attractive from a learning POV (point of view).
Thanks to Ava for explaining this to us.
I actually understand it.
I see problem especially in image recognition as it will be bias to a specific race of people since the training data is bias....
Lab 2 is about precisely that. Debiasing with CNNs
Really Great Learning and explanation.
Soo good lecture so clear appreciate you for this
autoencoder stands for "automatically encoding data"? I thought it was "self encoding"?
2. I don't think the VAE is inspired by the autoencoder. Instead, it has derived from variational methods and only happened to be autoencoding
3. the reparametrization noise is not drawn "from a prior distribution" but it's distribution just happens to coincide with the typically used prior
Thank you.
Thanks! 由衷感谢！
Hello Ava I'm from Brazil, can't wait to see too. I'm following the classes.
Where can I take my doubts from previous lessons?
you can join the what app group you can find the link in the second lecture comments
@@shivangagarwal6703, Can you please share the link? Thanks in advance!
Shouldn't the term be dz/dphi instead of df/dphi?
ahhh, can't wait to see this!!!!!
Amazing lecture !
just awesome explanations...
Where to get all lectures?
introtodeeplearning.com/ this is the course website.
@@victorsergio Thanks
Mind blowing!!!
Can anyone recommend me best online course for deep learning in python, i mean course should have detailed explanation in deep learning,specially in CNNs
you should check coursera deep learning by andrew yang, after that there is also tensorflow in practice, but NOTE:
- you should already by familiar with python
- you should know the deffrence between deep N, and machine learning so that you'll know what to learn
@@badreddineberrehal1624 Thanks a lot
great video!
The link for the database: www.dropbox.com/s/bp54q547mfg15ze/train_face.h5?dl=1 does not exist anymore. Cannot run the google collab code.
github.com/aamini/introtodeeplearning/issues/82
@@AAmini Thanks!
very nice
This is an outstanding lecture.
Great!
Thank you!
Amazing❤🖤🖤🖤🖤
7:48
Thanks
Two big reasons I watched the video.
Love.
Big
bitties
It's a lie, first face is mine. Or wait...
i want to be at MIT too :(
stuck at an IIT
Which IIT? That is what matters.
Jaha hamare sapne pure hote hai..waha aapka struggle shuru hota hai.....
@Demon King hahah! sai baat hai. pr dil maange more. Bahar ki univ k instructors bht achhe lagte, isliye bolra tha
Wierd flex, but okay.
@@amanbansiwal37 it's not much but it's honest work XD
Thank you very much Ava, excellent presentation. God bless you.