Daigo Fujiwara (@AcdFendder) - Perfil do Twitter

＜産経抄＞転覆事故、生徒の命奪った平和学習 sankei.com/article/202604… 近年、これほどはらわたが煮えくり返る話は珍しい。同志社国際高校が研修旅行のコースに、米軍普天間飛行場の名護市辺野古の移設工事現場の抗議船乗船を「平和学習」として組み込み、転覆事故で生徒を死なせた件である。

日本語

10

484

4.1K

61.4K

Daigo Fujiwara@AcdFendder·9h

真面目に金策を考えないとなぁと最近思っている

日本語

0

2

113

Daigo Fujiwara@AcdFendder·13h

うーん、不思議 VAEとかGANとかTransformerとかなんでもやが、それなりに哲学的にというか人間の認知からのアナロジーで上手く行きそうなのが納得できるんやがDiffusionだけはほんまに納得いってない。数式では分かるが

atulit@atulit_gaur

i love diffusion models and idk how but i went down a rabbit hole to trace their origins and honestly this stuff is kinda beautiful when you see how it evolved like it didn’t just appear out of nowhere… it’s literally years of ideas stacking on top of each other it starts with "deep unsupervised learning using nonequilibrium thermodynamics" (sohl-dickstein et al., 2015) -> slowly destroy data with noise -> learn to reverse it sounds simple but that’s the seed then comes "generative modeling by estimating gradients of the data distribution" (song and ermon, 2019) -> learn gradients of data distribution -> understand where real images live this is where score-based models come in then things get real with "denoising diffusion probabilistic models" (ho et al., 2020) -> clean probabilistic formulation -> stable training -> actually works this paper made diffusion practical then "diffusion models beat gans on image synthesis" (dhariwal and nichol, 2021) -> diffusion > gans this was a big shift but still slow as hell then the most important jump imo "high-resolution image synthesis with latent diffusion models" (rombach et al., 2021) -> compress images with vae -> run diffusion in latent space -> huge speedup this is literally the backbone of stable diffusion parallel to all this you have "learning transferable visual models from natural language supervision" (radford et al., 2021) -> align text and images -> enable text-guided generation that’s clip and finally stable diffusion (rombach et al., 2022 release implementation) -> vae + unet + clip -> trained on large-scale data like laion what’s crazy is the core idea never changed add noise -> remove noise -> repeat but layered with better math, better architectures, and better data also something about this feels kinda philosophical instead of solving a hard problem directly they broke it into tiny easy steps don’t generate an image just make it slightly less noisy each time again -> again -> again and somehow that builds something meaningful

日本語

0

6

781

Daigo Fujiwara retweetou

atulit@atulit_gaur·1d

i love diffusion models and idk how but i went down a rabbit hole to trace their origins and honestly this stuff is kinda beautiful when you see how it evolved like it didn’t just appear out of nowhere… it’s literally years of ideas stacking on top of each other it starts with "deep unsupervised learning using nonequilibrium thermodynamics" (sohl-dickstein et al., 2015) -> slowly destroy data with noise -> learn to reverse it sounds simple but that’s the seed then comes "generative modeling by estimating gradients of the data distribution" (song and ermon, 2019) -> learn gradients of data distribution -> understand where real images live this is where score-based models come in then things get real with "denoising diffusion probabilistic models" (ho et al., 2020) -> clean probabilistic formulation -> stable training -> actually works this paper made diffusion practical then "diffusion models beat gans on image synthesis" (dhariwal and nichol, 2021) -> diffusion > gans this was a big shift but still slow as hell then the most important jump imo "high-resolution image synthesis with latent diffusion models" (rombach et al., 2021) -> compress images with vae -> run diffusion in latent space -> huge speedup this is literally the backbone of stable diffusion parallel to all this you have "learning transferable visual models from natural language supervision" (radford et al., 2021) -> align text and images -> enable text-guided generation that’s clip and finally stable diffusion (rombach et al., 2022 release implementation) -> vae + unet + clip -> trained on large-scale data like laion what’s crazy is the core idea never changed add noise -> remove noise -> repeat but layered with better math, better architectures, and better data also something about this feels kinda philosophical instead of solving a hard problem directly they broke it into tiny easy steps don’t generate an image just make it slightly less noisy each time again -> again -> again and somehow that builds something meaningful

English

5

22

250

11.6K

Daigo Fujiwara@AcdFendder·13h

というか乳首が感じる必然性って生物的に無いから本来こっち用だったんじゃないか？？と思った。あとで調べよ(何の話)

ロンプラ@lonely_planet_

乳首が性感帯だったので、母乳育児は頭おかしくなるくらい気持ちよかったでもその気持ちよさは性的な気持ちよさではなく、オキシトシンがどばどば出て多幸感のなかでキマってる感じの気持ちよさチクイキできる人は母乳育児おすすめ✨ ちなみに妊娠中のチクオナは母乳出て楽しいよ！(早産リスクあり)

日本語

0

6

1.5K

Daigo Fujiwara retweetou

日本経済新聞電子版（日経電子版）@nikkei·16h

日本発「伊藤の公式」が支えるウォール街　莫大な富もたらす数学研究 nikkei.com/article/DGXZQO… 「ピタゴラスの定理に次いで応用された」（米科学アカデミー）──。故・伊藤清氏が編み出した数式は、デリバティブの価格予測から生物の絶滅確率まで幅広い分野で不可欠に。人類に富を生み出したのは「純粋な数学研究」でした。