To me, the critical and effective way of educating and enlightening is the step-by-step reasoning coupled with powerful animations. This video has certainly achieved that. Thanks so much!
I am sharing this and will endorse people in my contact to subscribe, as exaplaining this "not so common" topics, with this much ease is really a art and this efforts of yours deserves a great amount of respect and appreciation.
I can visualize autoencoders better now. Keep doing animations. My brain just encodes animation data easily And I need to decode them in exam paper / seminar
Nice explanation but I think two key aspects are missing (maybe planned to show up in later videos): 1. the connection to transformerts. 2. the fact that latent space allows you to make two models speek the same language (like the idea of CLIP and how its used in DallE)
Hi, thank you for the feedbacks ! Indeed these aspects are very important in modern architectures, but I feel like I would need to introduce a lot of other concepts to get there. It's definitely something I'll treat in future videos.
Thanks for the comment, in fact taking a simple interpolation is perfectly fine when your latent space is "in order". It should have some properties like being somewhat continuous, which is not imposed by a simple autoencoder. However VAEs do have such a latent space.
Could you make a video on common dimensionality reduction methods like PCA and projection (linear discrimants) etc? I’ve always been interested in when they should be applied but not the other. Anyways, nice video very underrated! Deserves more exposure! T^T
Thank you ! Yep that's the plan for the very next video: it will be an explanation of how several visualization methods work, there will probably be PCA, t-SNE and UMAP
great video. knew about encoders from the transformer model where the optimization criterion for embedding is the output of the decoder for the classification/generation task measured by eg. cross entropy loss and i know about word2vec where the optimization criterion is dot product similarity of co-occuring words. i did not know that in autoencoders the optimization criterion is minimizing the loss over reconstructing the original input. nice.
For the applications like Domain Adaptation and Image Colorization how does the loss function look like for an AutoEncoder ? Also you said that the MSE Loss is used but then in that case a trivial solution exists where the image is copied pixel by pixel and the Network Learns Nothing. How is that problem taken care of ?
@@rishidixit7939 Hi, I'm not familiar with those two tasks, but for Image Colorization an MSE would probably do just fine ? For preventing the Network to simply copy the image pixel by pixel, we have the bottleneck layer! Remember that this layer has a lot fewer neurons than there are pixels, so you can't just "copy" the values :)
Thank you ! Indeed the voiceover is generated by an AI, but it is my own voice that I cloned. I'm using Elevenlabs. Did that annoy you or got you out of the video ? :(
To me, the critical and effective way of educating and enlightening is the step-by-step reasoning coupled with powerful animations. This video has certainly achieved that. Thanks so much!
Thank you for your comment !
I have to argue that there is a fundamental difference between educating and enlightening.
Legendary algorithm pull. I love educational content like this. Road to 1M!
Thanks :)
this channel is a hidden gem!
Thank you !
Amazing how high quality your videos are. Hope you will have much more subscribers soon enough. This quality definetly deserve that.
I am sharing this and will endorse people in my contact to subscribe, as exaplaining this "not so common" topics, with this much ease is really a art and this efforts of yours deserves a great amount of respect and appreciation.
Wish Goodluck for this new channel
this explanation was exactly what I needed... I was having a hard time understanding the concept
I can visualize autoencoders better now. Keep doing animations.
My brain just encodes animation data easily
And I need to decode them in exam paper / seminar
Awesome video and animations bro. Its so amazing!! Keep doing more videos, I'll stay tuned!
Thank you, I'm not planning on stopping yet :)
Thank you very much for your videos. I am waiting for the next one about the VAE.
Thanks, hope I can post it this month :)
Excellent pace and choice of words. A video on UNET would be great
Awesome video! Thank you for your perfect explanation!!
Clear and concise explanations, awesome!
@@gabberwhacky Thanks
This channel is so underrated. Amazing explanation!
Thanks :)
Perfect tone and pace. Thanks.
Thanks !
wow, that was such a good video! Thanks for that
i just found your channel and fall in love with it. thank you !
Thanks for the kind words !
Thanks for making such an intuitive and insightful video! Cant wait for more content from this channel!
@@thmcass8027 Thanks !
Very good and easy yo understand content, i love when channels like yours make hard concept that easy to understand.
Thank you !
Awesome content.❤ The reasoning and intricate animation are mindblowing. Eagerly waiting for VAE video 😊
Thanks !
Nice explanation but I think two key aspects are missing (maybe planned to show up in later videos):
1. the connection to transformerts.
2. the fact that latent space allows you to make two models speek the same language (like the idea of CLIP and how its used in DallE)
Hi, thank you for the feedbacks ! Indeed these aspects are very important in modern architectures, but I feel like I would need to introduce a lot of other concepts to get there.
It's definitely something I'll treat in future videos.
This video is both informative and visually appealing. Thanks!
Many thanks :)
Thank you! You made it lucid.
Thank you for your comment !
Good job man! Nice graphical representations. Easy to follow.
Thank you so much !
Great video
Great video! Waiting for the one on VAEs and other topics
@@aryankashyap7194 Thanks, it will probably be up before the end of the summer :)
Great content, hope you can get more exposure!
Thanks :)
incredibly good content.keep up the good work!
Thank you !
Perfect animated and well explained. Thank you 👍 subscribed 😊
Thank you !
What would you do if you wanted to find a middle between two points in latent space if simple interpolation produces garbage results?
Thanks for the comment, in fact taking a simple interpolation is perfectly fine when your latent space is "in order".
It should have some properties like being somewhat continuous, which is not imposed by a simple autoencoder. However VAEs do have such a latent space.
Thanks for this wonderful content.
Thank you !
Could you make a video on common dimensionality reduction methods like PCA and projection (linear discrimants) etc? I’ve always been interested in when they should be applied but not the other. Anyways, nice video very underrated! Deserves more exposure! T^T
Thank you ! Yep that's the plan for the very next video: it will be an explanation of how several visualization methods work, there will probably be PCA, t-SNE and UMAP
great video. knew about encoders from the transformer model where the optimization criterion for embedding is the output of the decoder for the classification/generation task measured by eg. cross entropy loss and i know about word2vec where the optimization criterion is dot product similarity of co-occuring words. i did not know that in autoencoders the optimization criterion is minimizing the loss over reconstructing the original input. nice.
Thanks a lot !
Nice video, keep it up
@@stormaref Thanks !
Great video! Are you planning on releasing the code used for it?
Thank you ! Yes, I'll make a github page for the channel, I'll put the link in the description when it's done.
For the applications like Domain Adaptation and Image Colorization how does the loss function look like for an AutoEncoder ? Also you said that the MSE Loss is used but then in that case a trivial solution exists where the image is copied pixel by pixel and the Network Learns Nothing. How is that problem taken care of ?
@@rishidixit7939 Hi, I'm not familiar with those two tasks, but for Image Colorization an MSE would probably do just fine ?
For preventing the Network to simply copy the image pixel by pixel, we have the bottleneck layer! Remember that this layer has a lot fewer neurons than there are pixels, so you can't just "copy" the values :)
Great work!
Thank you !
Audio has latency around 4:09 with video!
Can you make a video on RNN and its variants?
Hi Sharjeel thanks for your comment !
RNN and other auto-regressive models are definitely on my to-do list. :)
8:02 Principal Component Analysis? 😉
or tsne/umap
4:10 Latent Space.
The GOAT
Please create more videos!
Sure will do aha
Do you use ai voiceover? Great video btw
Thank you ! Indeed the voiceover is generated by an AI, but it is my own voice that I cloned. I'm using Elevenlabs. Did that annoy you or got you out of the video ? :(
My name jeff.
Hi jeff