[HN Gopher] Diffusion models are evolutionary algorithms
___________________________________________________________________
Diffusion models are evolutionary algorithms
Author : che_shr_cat
Score : 52 points
Date : 2024-11-09 22:51 UTC (4 days ago)
(HTM) web link (gonzoml.substack.com)
(TXT) w3m dump (gonzoml.substack.com)
| adamnemecek wrote:
| They are all bialgebras.
| tbalsam wrote:
| I saw this comment, thought, "yeah this reminds me of Adam
| Nemeck", and lo and behold.
| will_byrd wrote:
| Interesting comment. Would you mind expanding on that
| observation? Are there any references you'd suggest looking at
| that help make the connection more clear? Thank you!
| bob1029 wrote:
| I have a hard time with the analogy due to how important
| population dynamics and solution diversity are to evolutionary
| algorithms.
|
| In an EA, each candidate in the population represents a complete
| potential solution. As the diversity & size of the population
| increases, the potential for convergence on high quality
| solutions also increases. I do not see the same concept in
| diffusion models.
| dawnofdusk wrote:
| >I do not see the same concept in diffusion models.
|
| I believe,
|
| Population size = noise dimensionality
|
| Population diversity = noise magnitude
| amelius wrote:
| But iiuc, an EA algorithm needs to keep an entire population
| in memory at once.
|
| I don't think this is the case for diffusion models.
| SubiculumCode wrote:
| In the end. Linear Regression.
| throwaway314155 wrote:
| Michael Levin's work is fascinating. Seems there's no field he
| can't help contribute to from a biological perspective.
___________________________________________________________________
(page generated 2024-11-13 23:00 UTC)