I would pay so much to have you as my teacher, that's not only the best video i've ever seen on deep leanring, but probably the most appealing way anyone ever taught me CS !
Thank you for creating such a nice explanation of VAE. When are you bringing the practical implementation video of VAE? please make a video for this as well.
Thanks, this video have many explanations that are missing from other tutorials on VAE. Like the part from 22:45 onwards. I saw a lot of other videos that didn't explain how the p and q functions were related to the encoder and decoder. (every other tutorial felt like they started talking about VAE, and then suddenly changed subject to talk about some distribution functions for no obvious reason).
great video ( am about half way through). I think at minute 13 there is a misstatement (I think). in general if g(x) is greater than h(x), if we find the max of h it doesn't mean we have located the max for g
Both VAE and AE map the input over a latent space, the difference lies on the **structure** of this latent space. The AE latent space is not "well-organized" as well as the VAE's latent space.
At the very end of the video Umar mentions that in "the next video" he will explain the coding of the VAE along with some examples. I could not find that video. Was it made? What is its title? Thanks 🙂
Latent variable is of low dimension compare to input which is of high dimension…so this low dimension latent variable contains features which are robust, meaning these robust features survive the encoding process coz encoding process removes redundant features….imagine a collection had images of cat and a bird image distribution, what an encoder can do in such a process is to outline a bird or cat by its outline without going into details of colours and texture….these outlines is more than enough to distinguish a bird from a cat without going into high dimensions of texture and colors
@@quonxinquonyi8570 that doesnt answer the question. Latent space in autoencoders dont capture semantic meaning , but when we enforce regularization on latent space and learn a distribution thats when it learns some manifold
@@prateekpatel6082 learning distribution means that you could generate from that distribution or in other words sample from such distribution…but since the “ sample generating distribution “ can be too hard to learn, so we go for reparametrization technique to learn the a standard normal distribution so that we can optimize
“ learning the manifold “ doesn’t make sense in the context of variational auto encoder….coz to learn the manifold, we try to approach the “score function” ….score function means the original input distribution….there we have to noised and denoised in order to get some sense of generating distribution….but the problem still holds in form of denominator of density of density function….so we use log of derivative of distribution to cancel out that constant denominator….then use the high school level first order derivative method to learn the noise by using the perturbed density function….
Great video on VAE Umar, but I keep with a question: At 11:49, the KL divergence is written as "D_{KL}({numerator}||{denominator}". While in 11:34 you wrote KL divergence as "D_{KL}{denominator}||{numerator}". Why is that? I guess it is not because of the "\minus" sign
🎯 Key points for quick navigation: 00:13 *understand fundamental ideas* 00:41 *explain autoencoder concept* 02:36 *issues with traditional autoencoders* 03:02 *introduce variational autoencoder* 04:13 *sampling from latent space* 25:02 *Model and prior choice.* 25:17 *Element-wise noise multiplication.* 25:32 *Learning log variance.* 26:00 *Maximizing ELBO for space learning.* 26:15 *Derivation of loss function.* Made with HARPA AI
Variational methods is a class of methods ive always struggled to understand. I think your explanation addresses some key questions but I thunk you jumped into the math too quickly. Questions im still wondering about: - what are the "variational" parameters? They are represented by phi, but what do they do/mean? - why is having the noise source necessary? What happens if i set epsilon to be 0 all the time?
Thanks for sharing . In the chicken and egg example, will p(x, z) be trackable? if x, z is unrelated, and z is a prior distribution, so p(x, z) can be writen in a formalized way?