Thank you for such a no math video. It's very rare to find videos with clear explanation for the intuition of the problem. Once we grab the idea, then the math seems more manageable. Thank you so much!
I'm always watching your video with gratitude. I studied basic image processing like CLAHE and segmentation like U-net. Nowdays, I'm studying RetinaNet that is one of object detection. But some concepts of RetinaNet like anchor boxes and transfer learning difficult to me. So, if you have free time, it would be nice upload a video about this object detection. I always feel grateful you. Thanks
Ohh finally I'm so glad you made tutorial on AAE please cover all aspects of AAE in image processing. Thanks so much you're best youtuber for image processing and deep learning. I'm your biggest fan just small request please take a bit more time in explaining code as I'm a biologist but interested in deep learning and image analysis. Thanks once again.
Well, the next video covers image generation where we generate MNOST images. AAEs are slow so to generate real images you will need lot more resources but the concepts I cover should definitely help.
Thank you for your great video! I saw a lot of notes which introduced too much about mathematical parts but ignored to tell why and how we need to use VAE. Your video helps me to understand why we need to learn a desirable distribution of the latent vector.
This is all great, I think my one quibble is that you are perhaps using a slightly nonstandard definition of "generative". Usually it means that we are modelling the distribution of the input space, and can therefore sample ("generate") new realistic inputs. For exactly the reasons you state, standard autoencoders don't do this, and therefore by definition are not generative models. Yes they can "generate" things but those things don't represent the input space and will probably be a "meaningless" mess. Whereas with variational autoencoders, they do model the input space and can therefore generate "realistic" inputs, so they are generative models.
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?
For what you are proposing you can even use 'standard' autoencoders.. Check this video ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-9zKuYvjFFS8.html