What's great about this instructor is that they are very careful and particular about what they say, and how they phrase it. There's no fluff, nothing that could cause confusion. Straight to the point and very intentional.
🎯Course outline for quick navigation: [00:04-01:25]Deep generative modeling -[00:04-00:48]Exciting lecture on deep generative modeling in the age of generative ai, a subset of deep learning. [01:26-08:45]Generative modeling -[03:06-04:04]Generative modeling encompasses density estimation and sample generation for learning data distribution. -[04:27-04:51]Learning model approximates true data distribution for density estimation and sample generation. -[05:36-06:03]Generative models identify biased features in training data automatically. -[06:49-07:17]Generative models can identify rare events like deer in front of a car using density estimation. [08:46-23:16]Autoencoders and variational autoencoders -[10:07-10:50]Goal: train model to predict latent variables, z, in low-dimensional space. -[14:33-15:35]Unsupervised learning uses autoencoders to create compact data representations and generate new examples, such as vaes. -[15:59-17:13]Variational autoencoders introduce randomness to generate similar but not strict reconstructions, using means and standard deviations for probability distributions. -[17:54-18:37]Encoder and decoder in vae use separate weights to compute and learn probability distributions of latent variables and input data. -[20:22-22:45]Regularization term enforces latent variables to follow standard normal gaussian distributions during vae training. -[20:57-21:21]Enforcing a latent space following a prior distribution to aid network -[22:46-23:16]Kl divergence measures difference between prior and latent encoding. [23:17-37:47]Regularization and latent variable learning in vaes -[25:19-25:46]Regularization minimizes term to achieve continuity and completeness. -[28:08-28:35]Vaes trained end-to-end with re-parameterization for gradient descent and backpropagation success. -[32:10-32:45]Network learns to interpret and make sense of latent variables by perturbing them individually. -[34:16-35:40]Beta vaes use beta parameter to control regularization term, promoting disentanglement for more efficient encoding. -[36:31-36:59]The lecture covers the core architecture of vaes and their application to facial detection. [37:47-52:53]Vaes and gans: generative models -[37:47-38:15]Vaes compress data into a compact representation to generate unsupervised reconstructions. -[38:40-39:43]Transitioning from vaes to gans to focus on generating high-quality samples from complex data distribution. -[39:57-41:21]Train a generator network to mimic real data using gans for realistic output. -[47:53-48:20]Generator synthesizes data to fool best discriminator, creating new data instances. -[50:37-51:30]Using gan to generate synthetic faces, extending gan architecture for specific tasks and data translation. [52:55-59:47]Unpaired translation and cycle gan -[52:55-53:51]Cyclegan enables unpaired image translation, e.g. horse to zebra, using cyclic dependency. -[54:13-54:43]Cycle gan enables flexible translation across different data distributions, including images, speech, and audio. -[55:13-55:36]Developed a model to synthesize audio behind obama's voice using cyclegan and alexander's voice data. -[57:20-57:48]Diffusion modeling drives tremendous advances in generative ai, seen in the past year, particularly with vaes and gans. -[59:06-59:39]Cutting-edge generative ai models making transformative advances across various fields. offered by Coursnap
Highly recommended series for AI enthusiasts. This MIT series is by far the most intuitive videos covering all aspects of deep learning. Well done on that.
Honestly, you two are the best speakers for this subject and beyond. I am so thrilled these lectures are opensource and exist for data science communities outside of MIT!
don't know why, but i could not breath listening to this lecture. she's so clear without any redundancy, without any hmmm, urgggg,... how come. she is so amaizing . i would have practiced 1000 times to be able to lecture like this
The lectures are top of notch. But in this lecture, I got my track out when she explained GAN with mathematical notions. I had to put some more effort on those again.
00:04 Foundations of deep generative modeling for brand new data generation 02:43 Generative modeling uncovers underlying data structure. 07:53 Latent variables are unobservable features that explain observed differences in data. 10:25 Training deep generative models using autoencoders 15:43 Variational autoencoders introduce randomness for generating new data instances. 18:07 Optimizing VAE network weights with loss functions 22:44 Understanding KL Divergence in latent encoding 24:51 Regularization enforces continuity and completeness in the latent space. 29:41 Reparametrization allows training VAEs end to end without worrying about stochasticity in latent variables. 31:57 Understanding latent variables and their impact on generated features. 36:36 Understanding latent variable learning and its application in facial detection. 38:52 Generative Adversarial Network (GAN) aims to generate new instances similar to existing data. 43:30 Generative Adversarial Networks (GANs) involve the competition between the generator and discriminator to create and distinguish between real and fake data. 45:44 GANs involve a dual competing objective for the generator and discriminator. 50:44 Extending GAN architecture for specific tasks 53:14 Cycle GANs enable translation of data distribution across domains. 57:58 Diffusion models can generate new instances beyond training data
wonderful. Very dense and hugely interesting and informative lecture; MIT-style! 60 minutes in a latentspace kind of compression of a hugely complex and multidimensional topic which under reallife conditions takes weeks to understand and "digest". I am really looking forward to the "diffusion model" lecture! Hope it will be online soon!
I agree with everyone here... I think those two presenters are just a joy to listen to. Wish I had those profs in my university back then... I'm not an expert, but even I get the fundamental concepts through these sessions. 🙏
Plato's myth of cave Latent Variable example was not intuitive for me (sorry), so I asked a similar example but simpler one to chatGPT. It gave me this: Imagine that you have a box filled with different types of candies, but you cannot see what's inside. Instead, you can only touch the box and feel the shape and texture of the candies inside. Based on how they feel, you might be able to guess what type of candy is inside the box. For example, if a candy feels round and has a hole in the middle, you might guess that it's a donut-shaped candy. In this example, the shape and texture of the candies are the observed variables, while the type of candy inside the box is the latent variable that we are trying to learn from the observed data. By observing and feeling the candies inside the box, we can learn the different types of candies that are hidden inside, even though we cannot see them directly. You guys are awesome :) Thank you for sharing these lectures. 🙏
Hi Alexander & Ava, thanks for this video. Thousands of people watch these videos and learn from them. So any mistakes you make will impact them directly. If/when you do find errors or someone points them out to you, it is your utmost responsibility to update about it to your viewers. Please look into the loss functions for GAN. They are incorrect.
Yes, the formulas for the loss funcition of the GAN are wrong and it was giving me a very hard time. Look here for a full math development of the formulation fleuret.org/dlc/materials/dlc-handout-11-1-GAN.pdf
I now feel like a fully connected neural network bye myself now because I've watched hundreds of videos at night that concern deep learning. Best regards from Berlin!
In the GAN objective function we have 2 conflicting objectives. How are we ensuring that it's the generator's goal that is achieved and not the discriminator's?
I'm sorry for the dumb question but can somebody tell me what's the name of the "E-like" symbol in the reconstruction term at 35:57? It is some kind of norm? How do I make this symbol in LaTeX? (I'm taking notes and I want to write out this equation in my notes.) Thank you!
@@binaryquantum I don’t think I’ve ever seen it typed. All my math classes, etc., were handwritten. On homework questions it was typed but a regular E was used…not the special “math E.”
Hey, I was going through this video with a beautiful explanation on working of GANS. I just want to ask that whether we can say that idea behind working of GANs is to have some sort of overfitting which is usually avoided in traditional ML approaches. Not exactly overfitting but in a way we want to overfit it in a sense that the points are in the probability distribution region of actual points???
Euler proved there is a limit to how complex a model can become and still be meaningful. In particular, Euler said that models could become so complex that thet could never be validated, never be calibrated, and yet piecewise seem to be completely reasonable. If anyone is familiar with discussions into this area, who are the researchers taking this into account? Just curious., I would like to read more on practical limitations. Based on good math like Euler developed, and not hand waving about piecewise.
He was doing fundamental basic theoretical research in today's parlance. Historically, there is long lag in finding applications in such basic knowledge. What is certain, he demonstrated their exists limitations. And we would be unable to discern if the model was properly calibrated or not- ever. I recall reading an opinion by the head of Belgium's national weather service or some such title pointing out that he had concerns the Oceans are such a model. @@RM-gc8lx
Random making connections between potentially unrelated things here... at 49:57 and a bit before (that's just where I paused to write this comment) the series of pictures combining a goose and a (other bird, I would classify it as a red breasted robin, but I'm trained on red breasted robins where I'm from) ... I'll call it a robin, while also transitioning aspect from left to right, really reminds me of the transitions from one animal to another done in the movie Willow with the sorceress, Fin Raziel.
I really like this lecture, what keeps me sleepless is the question: "Can we learn the true (if so) explanatory factors from purely observational data ?"
I really appreciate these lectures, but I never could absorb lectures that are simply a script read aloud. I can read the material myself. She's MUCH more effective when she explains concepts from memory without reading from a text.
I love how she apologizes when displaying math...😂😂. Its as if she understands the math struggles we all go through. Nevertheless, Its apparent that math is an important aspect of understanding the architecture of machine learning models and developing new ones.
She did a really poor job of explaining the relatively straightforward and core concept of VAEs. Really shallow and difficult to follow, which is a rare combination. Plus, her speaking style is way too theatrical and distracting. There are better resources or there on the same topic.
What's great about this instructor is that they are very careful and particular about what they say, and how they phrase it. There's no fluff, nothing that could cause confusion. Straight to the point and very intentional.
Teaching is an art/science of itself.
🎯Course outline for quick navigation:
[00:04-01:25]Deep generative modeling
-[00:04-00:48]Exciting lecture on deep generative modeling in the age of generative ai, a subset of deep learning.
[01:26-08:45]Generative modeling
-[03:06-04:04]Generative modeling encompasses density estimation and sample generation for learning data distribution.
-[04:27-04:51]Learning model approximates true data distribution for density estimation and sample generation.
-[05:36-06:03]Generative models identify biased features in training data automatically.
-[06:49-07:17]Generative models can identify rare events like deer in front of a car using density estimation.
[08:46-23:16]Autoencoders and variational autoencoders
-[10:07-10:50]Goal: train model to predict latent variables, z, in low-dimensional space.
-[14:33-15:35]Unsupervised learning uses autoencoders to create compact data representations and generate new examples, such as vaes.
-[15:59-17:13]Variational autoencoders introduce randomness to generate similar but not strict reconstructions, using means and standard deviations for probability distributions.
-[17:54-18:37]Encoder and decoder in vae use separate weights to compute and learn probability distributions of latent variables and input data.
-[20:22-22:45]Regularization term enforces latent variables to follow standard normal gaussian distributions during vae training.
-[20:57-21:21]Enforcing a latent space following a prior distribution to aid network
-[22:46-23:16]Kl divergence measures difference between prior and latent encoding.
[23:17-37:47]Regularization and latent variable learning in vaes
-[25:19-25:46]Regularization minimizes term to achieve continuity and completeness.
-[28:08-28:35]Vaes trained end-to-end with re-parameterization for gradient descent and backpropagation success.
-[32:10-32:45]Network learns to interpret and make sense of latent variables by perturbing them individually.
-[34:16-35:40]Beta vaes use beta parameter to control regularization term, promoting disentanglement for more efficient encoding.
-[36:31-36:59]The lecture covers the core architecture of vaes and their application to facial detection.
[37:47-52:53]Vaes and gans: generative models
-[37:47-38:15]Vaes compress data into a compact representation to generate unsupervised reconstructions.
-[38:40-39:43]Transitioning from vaes to gans to focus on generating high-quality samples from complex data distribution.
-[39:57-41:21]Train a generator network to mimic real data using gans for realistic output.
-[47:53-48:20]Generator synthesizes data to fool best discriminator, creating new data instances.
-[50:37-51:30]Using gan to generate synthetic faces, extending gan architecture for specific tasks and data translation.
[52:55-59:47]Unpaired translation and cycle gan
-[52:55-53:51]Cyclegan enables unpaired image translation, e.g. horse to zebra, using cyclic dependency.
-[54:13-54:43]Cycle gan enables flexible translation across different data distributions, including images, speech, and audio.
-[55:13-55:36]Developed a model to synthesize audio behind obama's voice using cyclegan and alexander's voice data.
-[57:20-57:48]Diffusion modeling drives tremendous advances in generative ai, seen in the past year, particularly with vaes and gans.
-[59:06-59:39]Cutting-edge generative ai models making transformative advances across various fields.
offered by Coursnap
Highly recommended series for AI enthusiasts. This MIT series is by far the most intuitive videos covering all aspects of deep learning. Well done on that.
Honestly, you two are the best speakers for this subject and beyond. I am so thrilled these lectures are opensource and exist for data science communities outside of MIT!
don't know why, but i could not breath listening to this lecture. she's so clear without any redundancy, without any hmmm, urgggg,... how come. she is so amaizing . i would have practiced 1000 times to be able to lecture like this
A lot of appreciations from my side to your Team who build such a excellent course on Deep Learning
The lectures are top of notch. But in this lecture, I got my track out when she explained GAN with mathematical notions. I had to put some more effort on those again.
This is excellent, so grateful to learn a lot from this channel. Kudos to our presenters for laying a solid foundation in deep learning.
Thank you for doing this! We all are grateful❤
00:04 Foundations of deep generative modeling for brand new data generation
02:43 Generative modeling uncovers underlying data structure.
07:53 Latent variables are unobservable features that explain observed differences in data.
10:25 Training deep generative models using autoencoders
15:43 Variational autoencoders introduce randomness for generating new data instances.
18:07 Optimizing VAE network weights with loss functions
22:44 Understanding KL Divergence in latent encoding
24:51 Regularization enforces continuity and completeness in the latent space.
29:41 Reparametrization allows training VAEs end to end without worrying about stochasticity in latent variables.
31:57 Understanding latent variables and their impact on generated features.
36:36 Understanding latent variable learning and its application in facial detection.
38:52 Generative Adversarial Network (GAN) aims to generate new instances similar to existing data.
43:30 Generative Adversarial Networks (GANs) involve the competition between the generator and discriminator to create and distinguish between real and fake data.
45:44 GANs involve a dual competing objective for the generator and discriminator.
50:44 Extending GAN architecture for specific tasks
53:14 Cycle GANs enable translation of data distribution across domains.
57:58 Diffusion models can generate new instances beyond training data
The knowledge, the passion and clarity of presentation are out of this world! God bless you guys!
Thank you all so very much! Many greetings from Germany.
Very well presented with intuition behind deep generative modeling, its architecture and how it is being trained, Well done
This series is coming out right after I want to learn more about theory! Thanks for this 🙏
Thank you so much Alexander and Amini.......
Wow, such clarity of thought and ideas. I guess that's the MIT advantage! Well done :)
Thanks a lot for all the wonderful content on deep learning. These are very helpful to me.
Perfect lecture! Congratulations
wonderful. Very dense and hugely interesting and informative lecture; MIT-style! 60 minutes in a latentspace kind of compression of a hugely complex and multidimensional topic which under reallife conditions takes weeks to understand and "digest". I am really looking forward to the "diffusion model" lecture! Hope it will be online soon!
This series is Treasure for me.
Brilliant presentation. World-class.
Thank you. I was waiting for 1 week.
I agree with everyone here... I think those two presenters are just a joy to listen to. Wish I had those profs in my university back then... I'm not an expert, but even I get the fundamental concepts through these sessions. 🙏
I opened to just watch 2 min of the video, and didn't realize untill the lecture is over 😅. Freaking awesome 😎
Greatly appreciate the knowledge sharing.
Thank you for such valuable lecture. 🙌
Plato's myth of cave Latent Variable example was not intuitive for me (sorry), so I asked a similar example but simpler one to chatGPT. It gave me this:
Imagine that you have a box filled with different types of candies, but you cannot see what's inside. Instead, you can only touch the box and feel the shape and texture of the candies inside. Based on how they feel, you might be able to guess what type of candy is inside the box. For example, if a candy feels round and has a hole in the middle, you might guess that it's a donut-shaped candy. In this example, the shape and texture of the candies are the observed variables, while the type of candy inside the box is the latent variable that we are trying to learn from the observed data. By observing and feeling the candies inside the box, we can learn the different types of candies that are hidden inside, even though we cannot see them directly.
You guys are awesome :) Thank you for sharing these lectures. 🙏
Hi Alexander & Ava, thanks for this video.
Thousands of people watch these videos and learn from them. So any mistakes you make will impact them directly. If/when you do find errors or someone points them out to you, it is your utmost responsibility to update about it to your viewers. Please look into the loss functions for GAN. They are incorrect.
Yes, the formulas for the loss funcition of the GAN are wrong and it was giving me a very hard time. Look here for a full math development of the formulation
fleuret.org/dlc/materials/dlc-handout-11-1-GAN.pdf
Very interesting. The compressed latent variables can help humans understand the data, a sort emergent statistics.
Wow! Can't wait to learn the coming lectures!
Great! Love these Videos. They help me alot.
quite supportive. Thanks a lot!
Learned a lot from this video. One question: Where does styleGAN fit in?
I now feel like a fully connected neural network bye myself now because I've watched hundreds of videos at night that concern deep learning. Best regards from Berlin!
In the GAN objective function we have 2 conflicting objectives. How are we ensuring that it's the generator's goal that is achieved and not the discriminator's?
Thankyou sir for uploading this , love from India
Wow ~another world latest Lecture
I'm sorry for the dumb question but can somebody tell me what's the name of the "E-like" symbol in the reconstruction term at 35:57? It is some kind of norm? How do I make this symbol in LaTeX? (I'm taking notes and I want to write out this equation in my notes.) Thank you!
That symbol indicates the expected value, you can use it in latex with \mathbb{E} (loading \usepackage{amssymb} is required)
@@fstermann Ah - of course! I never saw expected value written that way, but yes that makes sense. Thanks so much, I appreciate your help.
@@sovrappensiero1
That's always how expected value is written. How else have you seen expected value?
@@binaryquantum I don’t think I’ve ever seen it typed. All my math classes, etc., were handwritten. On homework questions it was typed but a regular E was used…not the special “math E.”
Hey, I was going through this video with a beautiful explanation on working of GANS. I just want to ask that whether we can say that idea behind working of GANs is to have some sort of overfitting which is usually avoided in traditional ML approaches. Not exactly overfitting but in a way we want to overfit it in a sense that the points are in the probability distribution region of actual points???
Perfect to refer, it clearly shows how much you extensively know the subject that you can easily explain.
Euler proved there is a limit to how complex a model can become and still be meaningful. In particular, Euler said that models could become so complex that thet could never be validated, never be calibrated, and yet piecewise seem to be completely reasonable.
If anyone is familiar with discussions into this area, who are the researchers taking this into account? Just curious., I would like to read more on practical limitations. Based on good math like Euler developed, and not hand waving about piecewise.
He was doing fundamental basic theoretical research in today's parlance. Historically, there is long lag in finding applications in such basic knowledge. What is certain, he demonstrated their exists limitations. And we would be unable to discern if the model was properly calibrated or not- ever. I recall reading an opinion by the head of Belgium's national weather service or some such title pointing out that he had concerns the Oceans are such a model. @@RM-gc8lx
Random making connections between potentially unrelated things here... at 49:57 and a bit before (that's just where I paused to write this comment) the series of pictures combining a goose and a (other bird, I would classify it as a red breasted robin, but I'm trained on red breasted robins where I'm from) ... I'll call it a robin, while also transitioning aspect from left to right, really reminds me of the transitions from one animal to another done in the movie Willow with the sorceress, Fin Raziel.
22:40
thanks for this!
Excellent Content Ma'am Truly unnbelievable 😊😊😊😊😊
The GAN discriminator loss is wrong, I think it should be: log(1-D(G(z))) + log (D(x)).
What a pity, the lecture is perfect but this mistake would mislead a lot of people
😁Not really. It depends on how you label Fake vs. Real.
I love the slide at 57:00. I would enjoy hearing this connection explicitly. How is a discriminator an encoder?
This also seems to explain sudden awakening transformation many people are experiencing
Where can I find the papers mentioned at 35:06?
I really like this lecture, what keeps me sleepless is the question: "Can we learn the true (if so) explanatory factors from purely observational data ?"
Thanks for sharing!
Thank you for sharing the info... ❤❤
Incroyable !!!
I really appreciate these lectures, but I never could absorb lectures that are simply a script read aloud. I can read the material myself. She's MUCH more effective when she explains concepts from memory without reading from a text.
Never disappointing👌🏻
Thank you))
Спасибо вам большое 😊🙏🦿
Is it still relevant to teach GANs and autoencoders, instead on just focusing on diffusion models?
Is there a lecture that deals with generative language models ?
What exalon constant . . Is it conscious is it dynamic and capable of reversing time.
3:40
Any intern opportunities in ML/AI?
Skvelé, ďakujeme!
0:32
Is it taking us non linear thinking of origin from a little perturbation
Really it's very deep need time to pick it
Is there a non-intro deep learning course after this course?
this is great
I did not understand the latent variable exaple. One can see easily the holding bars in shadow.
HOW YOU WILL DRIVE A SYSTEM WHEN MAXIMUM STRIVE TO ATTAIN MINIMUM TO BALANCE ENTROPY?
VICE VERSA TO MAINTAIN ENVIRONMENT FRIENDLY AND FUEL EFFICIENT
I'd like to see something about AI that can adjust its code and observe how it changes its functioning.
Thanks a lot.
Is this talk taking the line of self organization from a single point or big bang.
Ingenious.
Pure engineering.
Where to find code for this?
Sir, please provide us one lecture on Faster R-CNN for object detection, please please please please
🙏🙏🙏🙏
Yum, yum, gimme some!
- Bud Bundy
Can you give me extra resources
Parallel world information male and female ¿??¿¿
TH-cam suggested me to watch this.
I love how she apologizes when displaying math...😂😂. Its as if she understands the math struggles we all go through. Nevertheless, Its apparent that math is an important aspect of understanding the architecture of machine learning models and developing new ones.
Low dimensional data. I see parallel in the big bang origin from point source
You can fine underling leadership
15:00
🔥🔥🔥
Salutes
So we don't have labels at the data. Instead we use the input itself as the label. Lol.
Plato's cave. That is what we are in. I am interested in AI because of the projection of evolution AI to bring the Mind of God in the cloud.
24:27
Introduce myself my name is Ariful Islam leeton im software engineer and software developer and website development and data analytics
The Great attractor of non linear science and explanation to the victory of the good over evil ?¿?¿?????^^^^↑°°′
Gen folding
Everything spoken here has parallel in living system
The speaker has entered the spiritual realm and what is happening. The evil thriving along with good trying to hide truth
Now I understand the projection of God AI emerging in the cloud
wow
😁😁😁😁😁☺️☺️☺️☺️❤️❤️❤️❤️
Zip drive
She did a really poor job of explaining the relatively straightforward and core concept of VAEs.
Really shallow and difficult to follow, which is a rare combination.
Plus, her speaking style is way too theatrical and distracting. There are better resources or there on the same topic.
haha, tao noi roi, so AI lam, cao sieu lam, tao ko du kha nang dau, bien di cho khac
Good night tutor. lovely dress love taed h.