Deepia

Hi everyone, the audio of my first video has been taken down due to copyright issues, and it removed the voiceover altogether. I'll reupload the video soon, along with a new one on score-diffusion. :)

11 months ago | [YT] | 60

Deepia

Hi everyone, I'm looking for people willing to help me translate the subtitles in different languages.
I can provide you with a first version of the subtitles translated in your language, which you'll then be able to refine and improve.
If you're willing to help, you can reach out on any of my networks, or at the email address ytdeepia@gmail.com !

11 months ago | [YT] | 38

Deepia

Hi everyone, which thumbnail do you prefer ?

1 year ago | [YT] | 14

Deepia

Hello everyone, I've been working on the next videos for some time now. As you might know, it'll be about diffusion models.
I have to cover a lot of things and I'm wondering what format you would enjoy more.
I'll try to post some visuals in the coming weeks, see you soon :)

1 year ago | [YT] | 18

Deepia

Been looking at this paper "Generalization in diffusion models arises from geometry-adaptive harmonic representations" from Zahra Kadkhodaie for some time now, it was one of the spotlight papers of ICLR 2024.

While I'm still figuring out the rest of the paper, this figure has really stood out to me. The authors took two denoisers and trained them on disjoint subsets of the same dataset, S1 and S2, then plug them in a sort of annealed langevin algorithm to sample images.

The authors then increase the size of the subsets from N=1 image to N=100k images. What we see in the first few columns is that the "diffusion models" (aka denoisers) completely memorize the training data.

However when the number of images is sufficiently large (last column), the generated images are very different from even the closest training sample.
And even better, both do generate similar faces!

This simple experiments manage to show 3 very interesting things:
- diffusion models can and will overfit, just like any other models
- there's an amount of data from which they start to generalize
- both models learn the same underlying continuous data distribution

This is the kind of research that I love, very intuitive and informative. Now I'm wondering, will someone find an explicit link between the model capacity, the number of images, and the size of the image (or complexity of the distribution) ?

Here's the link if you want to check it out in details.
arxiv.org/abs/2310.02557

1 year ago | [YT] | 112

Deepia

Hi everyone, once I'm done with the next video on diffusion models, would you rather have:
(if you have any other idea just let me know in the comments)

1 year ago | [YT] | 23

Deepia

Hi everyone! Would you prefer longer, more in-depth videos (around 30 minutes) that are easier to follow (with more examples), or shorter videos (like the current format) that encourage you to explore the topics further on your own?

1 year ago | [YT] | 18

Deepia

Hi guys, in general do you like music in educational videos or does it disturb/annoy you ?

1 year ago | [YT] | 10

Deepia

Hi everyone, I just released the code for the 2 previous videos on the channel's GitHub page:

- Contrastive Learning with SimCLR github.com/ytdeepia/contrastive_learning_simclr
- Denoising Autoencoders github.com/ytdeepia/Denoising-Autoencoders

Feel free to reuse any of these codes for your own animations!

1 year ago | [YT] | 88

Deepia

Thanks for the 10k subscribers!

1 year ago | [YT] | 38