Wow that dog barking noise tripped my brain out so hard. Because my neighbor's dog always barks, my brain tuned out the sound of the bark until I reasoned he was taking about the sound of dogs barking. Neural networks aren't intelligent enough to behave in these ways.
Thanks for the vid, now I finally understand VAEs. I would also highly recommend watching the MIT Deep Generative Modelling video to better understand the technical details of VAEs.
Thanks for intuitive explanation. I'm really looking forward to see more detailed exploration on the VAE and its variants as noted in the last of the video.
Brilliant explanation! I have watched many videos on this topic, but most of them either throw some weird and unknown mathematical equation at you, which they just assume that you'll understand without a proper explanation and the rest just throws lines of python code at you, where the functions and parameters have thicc statistical names. You explained this like it is just a piece of cake! Thank you. :D
Thanks for this! Helped me understand the need for defining a region for these pools and consequently, having the K-L divergence in optimization. Up until now, I only looked at that regularization term as intentionally having information loss and now it makes sense that we need that to make the generator more useable for "varying" outputs.
Trying to make this as accessable as possible. It is a hard topic and sometimes I might hide behind that jargon. But I'll try to explain myself when I can
6:17 If passing in a random vector outputs garbage, then there are excess degrees of freedom in the vector. The variational autoencoder seems to be limiting the set of input vectors, so when we choose one from the limited set, we're assured it won't output garbage.
I think the reason the latent code is important is because that layer, that middle layer, has far fewer neurons than the input. So anything that's produced from there - has to come from a compressed form of the input.
QUESTION CONCERNING VAE! Using VAE with images, we currently start by compressing an image into the latent space and reconstructing from the latent space. QUESTION: What if we start with the photo of adult human, say a man or woman 25 years old (young adult) and we rebuild to an image of the same person but at a younger age, say man/woman at 14 years old (mid-teen). Do you see where I'm going with this? Can we create a VAE to make the face younger from 25 years (young adult) to 14 years (mid-teen)? In more general term, can VAE be used with non-identity function?
Hy, that was the best var Autoencoder video I found on the internet, so thanks a lot, it realy helped ! I have 2 questions regarding min 10:22 continious region. 1: (if i understood it correctly this is a no): is the number of dog-verctors in the dog pool equal to the number of dog pics in the training-set ? 2: if you take the most average dog-verctor from the d-pool, to make it short lets say: [70, 10, 0.4] than could the whole pool be descirbed as each of the values has it´s range like: [70(+/-10), 10(+/- 2, 0.4(+/- 0.02) ] and as long as all values from a new latent space vector are in this range, I am in the dog pool and therfore generate an okey-looking dog ? (little bonus question so the number of values in the vector and the range of each determines how much different dogs the network is able to create ?) thank you in advance, i hope my question was understandable
Hi, 1: You understood that correctly, so no. As the region is continiuous, it contains an infinite amount of vectors. On the other hand, you know only as many vectors of that region as you have input images (as you generate one for each image). 2: Not every dog image leads to a vector withing this pool and not every vector within this pool generates a dog image. This is due to the fact that a) we don't really understand how NNs function internally and b) these "pools" are just an explanation of what's wrong with traditional AE. That is, they do not have to really exist in the "real world". 3: As traditional AE decoders are deterministic, yes. If your latent vector can only have one value, you can only generate one image. The "range" shown in the video is a slight simplification of what is really going on. That is, you do not set hard bounds for your latent variables, but you formulate this as minimizing the KL-divergence (Kullback-Leibler-divergence, i.e. the "distance" of two distributions), so that the latent distribution does not strive away too much from the standard distribution. I hope my answers were both understandable and correct :)
really really good video. Could you tell me something about the Gaussian prior on the bottleneck. 1) Do we learn the parameters of this Gaussian? 2) Is it only 1 Gaussian, or as you said, it is really a mixture of Gaussians (mathematically speaking)? Thanks
At 7:12, you said that generative models need to learn these "pools" or distribution. Which part of the autoencoder is that? Or is it separate from that? To my understanding, the autoencoder alone just learns the weights for the encoder and decoder.
Understood nothing about how this model works. Oversimplifications and storytelling makes it unpaired with the how the real thing work. Now I know : AE is reducing the input data into a smaller vector, VAE can generate blurry image. What I don't know : What is happening to input data and the dataset, what this pool intuition is for?
I think is a very big mistake to say that auto encoders cannot used to generate data. That is very wrong and there are multiple applications which use images as inputs to generate images like for example how the baby from two parents will look like.
What happens if you know how many vector elements are needed to accurately define what you want to reproduce, and then add a few more that aren't defined by the input image, but represent the class of the desired output. Will this force all of the vector elements into their own pool? So you can pick any random vector and add to it the representation of the class, to only pick from that pool. This strategy works for the 'image painting' network by Andrej Karpathy, and it's how I switched between images for a different kind of image tweening. I still wonder exactly what kind of network the 'image painter' actually is. I'm guessing that the same technique should also work for a generative auto encoder. I came up with the idea based on how a person learns something; you get more than one input - I.e a picture and description, that goes in (both presented at the input, rather than one at the input and the other at the output), and is then mapped to just the wanted description.
awesome and simple explanation. I was confused and wondering about the sampling part that VAE's do because i didn't understand what was meant by sampling a latent vector from a distribution. But you made it so easy to understand. Thanks a lot. Keep up the good work
Nice video! I have two questions: You show that the pool of the VAE is continuous, but it also shows blanks, eg. all space isnt covered by the numbers. What does a sampling from these regions gives? Is it still close to a number? Second question, does the size of the pool affect the quality of an image generated? Like giving more space to the VAE allows it to learn with less constrains? Thanks!
@@CodeEmporium thanks for the reply AJ. I was really surprised the way you changed your accent in such a short span of time, at one point I couldn't believe the fact that you're Thamil. Your content is amazing, I don't want to be selfish here, but I'd be happy if you can do more video's on NLP.
Well, that's actually a totally wrong conceptual explaination of a VAE. Moreover, in the video you didn't name some absolutely cruicial points about VAEs, that one would expect to hear. Moremoremoreover, there are plenty of statistical and mathematical things, that are not obvious at all and need to be explained when speaking about VAEs. So this is indeed an explaination, but quite a bad one I could be more specific if anybody is interested, so let's start some discussion in the comments :D
on the final slide, how did you find the latent vectors for the VAE that generate images similar to the images generated by GAN? or were the images on the right the result of encoding and decoding the images generated by the GAN using the VAE? then the VAE seems really bad at its original job