Continue the discussion on The Sound of AI Slack Community -> valeriovelardo.com/the-sound-of-ai-community/
@ValerioVelardoTheSoundofAI
Жыл бұрын
I made a mistake in the video, claiming that "AudioLM = w2v-BERT+ MuLan", when the correct architecture is "AudioLM = w2v-BERT+ SoundStream".
@musicish20
Жыл бұрын
Fantastic explanation of the model/paper! Thank you for this. I also agree with all of your points at the end - ai applications such as this need to be open-sourced to prevent the (seemingly) rapid creation of monopolies in this space.
@user-uj8qn9ed3c
2 ай бұрын
great explanation and already join to your slack community 🔥🔥
@dizzler-beats
Жыл бұрын
great insights on how these models work, thank you !
@CalendulaF
Жыл бұрын
Impressive achievement by google, but also valid criticism by you! Thanks for pointing out the issues not only about the limitations of this particular model but also about the practice of communicating only the pieces necessary to impress a broader audience (and perhaps shareholders and investors).
@ValerioVelardoTheSoundofAI
Жыл бұрын
Nice way of putting it - I may steal it in the future :D
@kamalakannank2741
6 ай бұрын
Thank you so much sir 🥹🥹 I have taken this paper for my final year project and stuck in these too much technical stuffs.. Thank god finally I found this video ❤🥰🥰
@mandilquioxtenlp1202
Жыл бұрын
Thank you for the great explanation!
@ValerioVelardoTheSoundofAI
Жыл бұрын
You're welcome!
@Sawaedo
Жыл бұрын
Thank you Valerio for your explanation and analysis. The paper will appear too impressive but the limitations that you exposed were necessary to understand the big picture.
@ValerioVelardoTheSoundofAI
Жыл бұрын
I second that the results are impressive. However, there are some fundamental limitations one should be aware of.
@musicai9043
Жыл бұрын
i agree, we should move into open sourcing if we want things to progress faster
@damiansaez5586
Жыл бұрын
Thank you Valerio for the detailed explanation!
@ValerioVelardoTheSoundofAI
Жыл бұрын
You're welcome Damian!
@JacobKinsley
Жыл бұрын
I'm gonna be honest with this ai model and say the music it generates is like the music equivalent of cheap vanilla ice cream with tiny chunks of dog food strewn throughout OK but just to be less rude, the best way I can explain it is that music generation ai is going through something similar to image generation ai. This is like when nvidia made that landscape drawing ai. It was novel, it was cool, and most of all it was incredibly realistic, but also couldn't do anything without it either being plain, or a straight up garbage output that doesn't look (or in this case sound) like anything particularly interesting.
@ValerioVelardoTheSoundofAI
Жыл бұрын
Fair point.
@matthewpublikum3114
Жыл бұрын
One thing though, it seems like the main point of these models are to provide a way of exploring latent spaces using text input rather than disentangling the variables through explicit formulation like VAEs.
@ValerioVelardoTheSoundofAI
Жыл бұрын
I would agree with your statement. In a sense, VAEs are less black boxy.
@jeremyuzan1169
3 ай бұрын
Great video
@AdrianFernandezFazio
Жыл бұрын
Great explanation Valerio! I agree about the long examples. Seems like it's having problems with the music forms . Greetings!
@ValerioVelardoTheSoundofAI
Жыл бұрын
Yep, the long-term structure is problematic.
@musiceverywhere3168
Жыл бұрын
You're the best!
@psychurch
Жыл бұрын
I have an idea of how to create a more robust semantic structure, would love to have a discussion. Slack community doesn’t seem to exist anymore and discord is not allowing any conversations?
@ValerioVelardoTheSoundofAI
Жыл бұрын
The Slack community exists. Are you having problems joining via the link on my website? 2 people have joined earlier today. Sometimes, I have to update the Slack invite link. Let me know...
@benediktzoennchen
Жыл бұрын
Fantastic video. Thank you for your criticism! The term "AI oligopoly" is precisely what I feel is on the horizon, but I could not name it so accurately.
@ValerioVelardoTheSoundofAI
Жыл бұрын
I'm also afraid we're headed that way :(
@psychurch
Жыл бұрын
Music + Language = MuLan
@TileBitan
Жыл бұрын
I understand where you come from, but I think you were too nitpicky. Sound ml is nothing compared to text and image generation, in the eyes of the public. The more models we have (and google's musicLM is producing incredibly good music compared to most previous attempts), the better. Get this branch of AI the attention it deserves Also disagree on your stance against big data models and in favour of artificially inducing information through symbolic structures. The point of AI is for machines to learn, and the ultimate goal is to feed terabytes of data and getting something akin to real intelligence. AI is not about putting constraints into the model, that's just regular machine learning and although is respectable, shouldn't replace these more brute force approaches. They should release their model like the stable diffusion guys, on that I agree 100%, but it is not our call
@ValerioVelardoTheSoundofAI
Жыл бұрын
Thank you for your perspective. I disagree that this model is "producing incredibly good music compared to most previous attempts". There are several models, also old ones, in the symbolic representation that are far superior from a compositional standpoint than MusicLM, showing coherent long-term structures. From my experience, the most successful attempts at generative music are those which mix symbolic knowledge and DL. My bet is on neuro-symbolic integration.
@TileBitan
Жыл бұрын
@@ValerioVelardoTheSoundofAI but it is not the same playing field. Researchers introduced artificial contraints to a model that wouldnt have picked up any long term coherent behavior, while musicLM is more about getting there without such constraints, just data and the right approach. To this i will say though, they mixed two approaches and one of them is more similar to what you describe, which may be the reason it outperformed expectations
@ValerioVelardoTheSoundofAI
Жыл бұрын
@@TileBitan fair point. That said, I think brute-forcing music generation through massive datasets / models isn't the most elegant, neither the most effective solution, and certainly isn't how human composers work. I like the connectionist approach, but I believe it has its limitations. Symbolic reasoning is a fundamental property of intelligence (mind).
Пікірлер: 33