Generative Multimodal Models are In-Context Learners

Beijing Academy of Artificial Intelligence

Tsinghua University

Peking University


Abstractย 

The human ability to easily solve multimodal tasks in context (i.e., with only a few demonstrations or simple instructions), is what current multimodal systems have largely struggled to imitate. In this work, we demonstrate that the task-agnostic in-context learning capabilities of large multimodal models can be significantly enhanced by effective scaling-up. We introduce Emu2, a generative multimodal model with 37 billion parameters, trained on large-scale multimodal sequences with a unified autoregressive objective. Emu2 exhibits strong multimodal in-context learning abilities, even emerging to solve tasks that require on-the-fly reasoning, such as visual prompting and object-grounded generation. The model sets a new record on multiple multimodal understanding tasks in few-shot settings. When instruction-tuned to follow specific instructions, Emu2 further achieves new state-of-the-art on challenging tasks such as question answering benchmarks for large multimodal models and open-ended subject-driven generation. These achievements demonstrate that Emu2 can serve as a base model and general-purpose interface for a wide range of multimodal tasks. Code and models are publicly available to facilitate future research.


ํ˜„์žฌ ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ ์‹œ์Šคํ…œ๋“ค์ด ์ฃผ๋กœ ๋ชจ๋ฐฉํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ๋Š” ๊ฒƒ์€ ์ธ๊ฐ„์ด ๋ฌธ๋งฅ ์†์—์„œ ์ ์€ ์ˆ˜์˜ ์‹œ๋ฒ”์ด๋‚˜ ๊ฐ„๋‹จํ•œ ์ง€์‹œ๋งŒ์œผ๋กœ ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ ์ž‘์—…์„ ์‰ฝ๊ฒŒ ํ•ด๊ฒฐํ•˜๋Š” ๋Šฅ๋ ฅ์ž…๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ, ์šฐ๋ฆฌ๋Š” ํฐ ๊ทœ๋ชจ์˜ ๋‹ค๋ชจ๋‹ฌ ๋ชจ๋ธ์˜ ๋งฅ๋ฝ ๋‚ด์—์„œ์˜ ๊ณผ์ œ-๋ฌด๊ด€ ํ•™์Šต ๋Šฅ๋ ฅ์ด ํšจ๊ณผ์ ์ธ ํ™•์žฅ์„ ํ†ตํ•ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ํ†ตํ•ฉ ์ž๋™ ํšŒ๊ท€ ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง„ ๋Œ€๊ทœ๋ชจ ๋‹ค๋ชจ๋‹ฌ ์‹œํ€€์Šค๋กœ ํ›ˆ๋ จ๋œ 370์–ต ๊ฐœ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๊ฐ€์ง„ ์ƒ์„ฑ์  ๋‹ค๋ชจ๋‹ฌ ๋ชจ๋ธ์ธ Emu2๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Emu2๋Š” ๊ฐ•๋ ฅํ•œ ๋‹ค๋ชจ๋‹ฌ ๋งฅ๋ฝ ๋‚ด ํ•™์Šต ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์‹œ๊ฐ์  ํ”„๋กฌํ”„ํŒ… ๋ฐ ๊ฐ์ฒด ๊ธฐ๋ฐ˜ ์ƒ์„ฑ๊ณผ ๊ฐ™์ด ์ฆ‰์„ ์ถ”๋ก ์ด ํ•„์š”ํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ์—๋„ ๋‚˜ํƒ€๋‚ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๋ช‡ ๊ฐœ์˜ ์˜ˆ์ œ ์„ค์ •์—์„œ ๋‹ค์–‘ํ•œ ๋‹ค๋ชจ๋‹ฌ ์ดํ•ด ๊ณผ์ œ์— ์ƒˆ๋กœ์šด ๊ธฐ๋ก์„ ์„ธ์›๋‹ˆ๋‹ค. ํŠน์ • ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด๋„๋ก ์ง€์‹œ ํŠœ๋‹๋˜์—ˆ์„ ๋•Œ, Emu2๋Š” ๋Œ€๊ทœ๋ชจ ๋‹ค๋ชจ๋‹ฌ ๋ชจ๋ธ๊ณผ ๊ฐœ๋ฐฉํ˜• ์ฃผ์ œ ๊ธฐ๋ฐ˜ ์ƒ์„ฑ์„ ์œ„ํ•œ ์งˆ๋ฌธ ๋Œ€๋‹ต ๋ฒค์น˜๋งˆํฌ์™€ ๊ฐ™์€ ๋„์ „์ ์ธ ๊ณผ์ œ์—์„œ ์ƒˆ๋กœ์šด ์ตœ๊ณ  ๊ธฐ๋ก์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ์„ฑ๊ณผ๋“ค์€ Emu2๊ฐ€ ๋‹ค์–‘ํ•œ ๋‹ค๋ชจ๋‹ฌ ๊ณผ์ œ์— ๋Œ€ํ•œ ๋ฒ ์ด์Šค ๋ชจ๋ธ ๋ฐ ์ผ๋ฐ˜์ ์ธ ์ธํ„ฐํŽ˜์ด์Šค๋กœ์„œ ๊ธฐ๋Šฅํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ฝ”๋“œ ๋ฐ ๋ชจ๋ธ์€ ํ–ฅํ›„ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ๊ณต๊ฐœ์ ์œผ๋กœ ์ด์šฉ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.





https://arxiv.org/abs/2312.13286