Latest AI news 2023/12/29 ์ตœ์‹  AI ๋‰ด์Šค


Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision, Language, Audio, and Action
Unified-IO 2: ์‹œ๊ฐ, ์–ธ์–ด, ์˜ค๋””์˜ค, ํ–‰๋™์„ ํฌํ•จํ•œ ์ž๊ธฐํšŒ๊ท€ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ชจ๋ธ ํ™•์žฅ


Allen Institute for AI

University of Illinois Urbana-Champaignย 

University of Washington


3์ค„์š”์•ฝ

1. "Unified-IO 2"๋Š” ์ด๋ฏธ์ง€, ํ…์ŠคํŠธ, ์˜ค๋””์˜ค, ํ–‰๋™์„ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์ตœ์ดˆ์˜ ์ž๊ธฐํšŒ๊ท€ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

2. ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ๋ฅผ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด, ์ด ๋ชจ๋ธ์€ ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ์„ ๊ณต์œ  ์˜๋ฏธ ๊ณต๊ฐ„์œผ๋กœ ํ† ํฐํ™”ํ•˜์—ฌ ๋‹จ์ผ ์ธ์ฝ”๋”-๋””์ฝ”๋” ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค.

3. "Unified-IO 2"๋Š” GRIT ๋ฒค์น˜๋งˆํฌ์—์„œ ์ตœ๊ณ ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ณ , ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ฐ ์ดํ•ด, ์ž์—ฐ์–ด ์ดํ•ด, ๋น„๋””์˜ค ๋ฐ ์˜ค๋””์˜ค ์ดํ•ด, ๋กœ๋ด‡ ์กฐ์ž‘ ๋“ฑ 35๊ฐœ ์ด์ƒ์˜ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ฐ•๋ ฅํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.


Abstractย 

We present Unified-IO 2, the first autoregressive multimodal model that is capable of understanding and generating image, text, audio, and action. To unify different modalities, we tokenize inputs and outputs -- images, text, audio, action, bounding boxes, etc., into a shared semantic space and then process them with a single encoder-decoder transformer model. Since training with such diverse modalities is challenging, we propose various architectural improvements to stabilize model training. We train our model from scratch on a large multimodal pre-training corpus from diverse sources with a multimodal mixture of denoisers objective. To learn an expansive set of skills, such as following multimodal instructions, we construct and finetune on an ensemble of 120 datasets with prompts and augmentations. With a single unified model, Unified-IO 2 achieves state-of-the-art performance on the GRIT benchmark and strong results in more than 35 benchmarks, including image generation and understanding, natural language understanding, video and audio understanding, and robotic manipulation. We release all our models to the research community.


ํ†ตํ•ฉ-IO 2๋Š” ์ด๋ฏธ์ง€, ํ…์ŠคํŠธ, ์˜ค๋””์˜ค, ํ–‰๋™์„ ์ดํ•ดํ•˜๊ณ  ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ์ตœ์ดˆ์˜ ์ž๊ธฐํšŒ๊ท€ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ๋ฅผ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด, ์ด๋ฏธ์ง€, ํ…์ŠคํŠธ, ์˜ค๋””์˜ค, ํ–‰๋™, ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค ๋“ฑ์„ ๊ณต์œ  ์˜๋ฏธ ๊ณต๊ฐ„์œผ๋กœ ํ† ํฐํ™”ํ•œ ํ›„ ๋‹จ์ผ ์ธ์ฝ”๋”-๋””์ฝ”๋” ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ๋„์ „์ ์ด๊ธฐ ๋•Œ๋ฌธ์—, ์šฐ๋ฆฌ๋Š” ๋ชจ๋ธ ํ›ˆ๋ จ์„ ์•ˆ์ •ํ™”์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์—ฌ๋Ÿฌ ๊ตฌ์กฐ์  ๊ฐœ์„ ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง„ ๋Œ€๊ทœ๋ชจ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์‚ฌ์ „ ํ›ˆ๋ จ ์ฝ”ํผ์Šค์—์„œ ์ฒ˜์Œ๋ถ€ํ„ฐ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด๋Š” ๋“ฑ์˜ ๊ด‘๋ฒ”์œ„ํ•œ ๊ธฐ์ˆ ์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด, ์šฐ๋ฆฌ๋Š” ํ”„๋กฌํ”„ํŠธ์™€ ๋ณด๊ฐ•์ด ํฌํ•จ๋œ 120๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹ ์•™์ƒ๋ธ”์„ ๊ตฌ์ถ•ํ•˜๊ณ  ๋ฏธ์„ธ ์กฐ์ •์„ ํ•ฉ๋‹ˆ๋‹ค. ๋‹จ์ผ ํ†ตํ•ฉ ๋ชจ๋ธ๋กœ, ํ†ตํ•ฉ-IO 2๋Š” GRIT ๋ฒค์น˜๋งˆํฌ์—์„œ ์ตœ๊ณ ์˜ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ณ  ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ฐ ์ดํ•ด, ์ž์—ฐ์–ด ์ดํ•ด, ๋น„๋””์˜ค ๋ฐ ์˜ค๋””์˜ค ์ดํ•ด, ๋กœ๋ด‡ ์กฐ์ž‘ ๋“ฑ 35๊ฐœ ์ด์ƒ์˜ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ฐ•๋ ฅํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ๋ชจ๋ธ์„ ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.

24b0d121e09c28a8699fe8b115ef0468e3a8eabb63

24b0d121e09c28a8699fe8b115ef046c62f72b4a9d

24b0d121e09c28a8699fe8b115ef046f584b9f9caa

24b0d121e09c28a8699fe8b115ef0468e5aee3b8

24b0d121e09c28a8699fe8b115ef0469963ff4bb

24b0d121e09c28a8699fe8b115ef046a7a64eece

https://arxiv.org/pdf/2312.17172.pdf