Vision-language foundational fashions are constructed on the premise of a single pre-training adopted by subsequent adaptation to a number of downstream duties. Two foremost and disjoint coaching eventualities are well-liked: a CLIP-style contrastive studying and next-token prediction. Contrastive studying trains the mannequin to foretell if image-text pairs appropriately match, successfully constructing visible and textual content representations for the corresponding picture and textual content inputs, whereas next-token prediction predicts the probably subsequent textual content token in a sequence, thus studying to generate textual content, in response to the required job. Contrastive studying allows image-text and text-image retrieval duties, reminiscent of discovering the picture that finest matches a sure description, and next-token studying allows text-generative duties, reminiscent of Image Captioning and Visual Question Answering (VQA). While each approaches have demonstrated highly effective outcomes, when a mannequin is pre-trained contrastively, it sometimes doesn’t fare effectively on text-generative duties and vice-versa. Furthermore, adaptation to different duties is commonly performed with advanced or inefficient strategies. For instance, so as to lengthen a vision-language mannequin to movies, some fashions must do inference for every video body individually. This limits the dimensions of the movies that may be processed to only some frames and doesn’t absolutely benefit from movement data accessible throughout frames.
Motivated by this, we current “A Simple Architecture for Joint Learning for MultiModal Tasks”, referred to as MaMMUT, which is ready to prepare collectively for these competing goals and which supplies a basis for a lot of vision-language duties both straight or through easy adaptation. MaMMUT is a compact, 2B-parameter multimodal mannequin that trains throughout contrastive, textual content generative, and localization-aware goals. It consists of a single picture encoder and a textual content decoder, which permits for a direct reuse of each elements. Furthermore, a simple adaptation to video-text duties requires solely utilizing the picture encoder as soon as and might deal with many extra frames than prior work. In line with latest language fashions (e.g., PaLM, GLaM, GPT3), our structure makes use of a decoder-only textual content mannequin and could be regarded as a easy extension of language fashions. While modest in measurement, our mannequin outperforms the state-of-the-art or achieves aggressive efficiency on image-text and text-image retrieval, video query answering (VideoQA), video captioning, open-vocabulary detection, and VQA.
 |
 |
 |
| The MaMMUT mannequin allows a variety of duties reminiscent of image-text/text-image retrieval (prime left and prime proper), VQA (center left), open-vocabulary detection (center proper), and VideoQA (backside). |
Decoder-only mannequin structure
One stunning discovering is {that a} single language-decoder is ample for all these duties, which obviates the necessity for each advanced constructs and coaching procedures offered earlier than. For instance, our mannequin (offered to the left within the determine under) consists of a single visible encoder and single text-decoder, related through cross consideration, and trains concurrently on each contrastive and text-generative varieties of losses. Comparatively, prior work is both not in a position to deal with image-text retrieval duties, or applies just some losses to just some components of the mannequin. To allow multimodal duties and absolutely benefit from the decoder-only mannequin, we have to collectively prepare each contrastive losses and text-generative captioning-like losses.
 |
| MaMMUT structure (left) is a straightforward assemble consisting of a single imaginative and prescient encoder and a single textual content decoder. Compared to different well-liked vision-language fashions — e.g., PaLI (center) and ALBEF, CoCa (proper) — it trains collectively and effectively for a number of vision-language duties, with each contrastive and text-generative losses, absolutely sharing the weights between the duties. |
Decoder two-pass studying
Decoder-only fashions for language studying present clear benefits in efficiency with smaller mannequin measurement (nearly half the parameters). The foremost problem for making use of them to multimodal settings is to unify the contrastive studying (which makes use of unconditional sequence-level illustration) with captioning (which optimizes the chance of a token conditioned on the earlier tokens). We suggest a two-pass strategy to collectively be taught these two conflicting varieties of textual content representations inside the decoder. During the primary go, we make the most of cross consideration and causal masking to be taught the caption era job — the textual content options can attend to the picture options and predict the tokens in sequence. On the second go, we disable the cross-attention and causal masking to be taught the contrastive job. The textual content options won’t see the picture options however can attend bidirectionally to all textual content tokens directly to supply the ultimate text-based illustration. Completing this two-pass strategy inside the identical decoder permits for accommodating each varieties of duties that had been beforehand arduous to reconcile. While easy, we present that this mannequin structure is ready to present a basis for a number of multimodal duties.
 |
| MaMMUT decoder-only two-pass studying allows each contrastive and generative studying paths by the identical mannequin. |
Another benefit of our structure is that, since it’s skilled for these disjoint duties, it may be seamlessly utilized to a number of functions reminiscent of image-text and text-image retrieval, VQA, and captioning.
Moreover, MaMMUT simply adapts to video-language duties. Previous approaches used a imaginative and prescient encoder to course of every body individually, which required making use of it a number of instances. This is sluggish and restricts the variety of frames the mannequin can deal with, sometimes to solely 6–8. With MaMMUT, we use sparse video tubes for light-weight adaptation straight through the spatio-temporal data from the video. Furthermore, adapting the mannequin to Open-Vocabulary Detection is completed by merely coaching to detect bounding-boxes through an object-detection head.
 |
| Adaptation of the MaMMUT structure to video duties (left) is straightforward and absolutely reuses the mannequin. This is completed by producing a video “tubes” characteristic illustration, much like picture patches, which can be projected to decrease dimensional tokens and run via the imaginative and prescient encoder. Unlike prior approaches (proper) that must run a number of particular person photos via the imaginative and prescient encoder, we use it solely as soon as. |
Results
Our mannequin achieves wonderful zero-shot outcomes on image-text and text-image retrieval with none adaptation, outperforming all earlier state-of-the-art fashions. The outcomes on VQA are aggressive with state-of-the-art outcomes, that are achieved by a lot bigger fashions. The PaLI mannequin (17B parameters) and the Flamingo mannequin (80B) have one of the best efficiency on the VQA2.0 dataset, however MaMMUT (2B) has the identical accuracy because the 15B PaLI.
 |
 |
| MaMMUT outperforms the state-of-the-art (SOTA) on Zero-Shot Image-Text (I2T) and Text-Image (T2I) retrieval on each MS-COCO (prime) and Flickr (backside) benchmarks. |
 |
| Performance on the VQA2.0 dataset is aggressive however doesn’t outperform giant fashions reminiscent of Flamingo-80B and PalI-17B. Performance is evaluated within the more difficult open-ended textual content era setting. |
MaMMUT additionally outperforms the state-of-the-art on VideoQA, as proven under on the MSRVTT-QA and MSVD-QA datasets. Note that we outperform a lot larger fashions reminiscent of Flamingo, which is particularly designed for picture+video pre-training and is pre-trained with each image-text and video-text knowledge.
 |
 |
| MaMMUT outperforms the SOTA fashions on VideoQA duties (MSRVTT-QA dataset, prime, MSVD-QA dataset, backside), outperforming a lot bigger fashions, e.g., the 5B GIT2 or Flamingo, which makes use of 80B parameters and is pre-trained for each image-language and vision-language duties. |
Our outcomes outperform the state-of-the-art on open-vocabulary detection fine-tuning as can also be proven under.
Key components
We present that joint coaching of each contrastive and text-generative goals is just not a straightforward job, and in our ablations we discover that these duties are served higher by totally different design selections. We see that fewer cross-attention connections are higher for retrieval duties, however extra are most well-liked by VQA duties. Yet, whereas this exhibits that our mannequin’s design selections may be suboptimal for particular person duties, our mannequin is more practical than extra advanced, or bigger, fashions.
 |
 |
| Ablation research displaying that fewer cross-attention connections (1-2) are higher for retrieval duties (prime), whereas extra connections favor text-generative duties reminiscent of VQA (backside). |
Conclusion
We offered MaMMUT, a easy and compact vision-encoder language-decoder mannequin that collectively trains a lot of conflicting goals to reconcile contrastive-like and text-generative duties. Our mannequin additionally serves as a basis for a lot of extra vision-language duties, attaining state-of-the-art or aggressive efficiency on image-text and text-image retrieval, videoQA, video captioning, open-vocabulary detection and VQA. We hope it may be additional used for extra multimodal functions.
Acknowledgements
The work described is co-authored by: Weicheng Kuo, AJ Piergiovanni, Dahun Kim, Xiyang Luo, Ben Caine, Wei Li, Abhijit Ogale, Luowei Zhou, Andrew Dai, Zhifeng Chen, Claire Cui, and Anelia Angelova. We want to thank Mojtaba Seyedhosseini, Vijay Vasudevan, Priya Goyal, Jiahui Yu, Zirui Wang, Yonghui Wu, Runze Li, Jie Mei, Radu Soricut, Qingqing Huang, Andy Ly, Nan Du, Yuxin Wu, Tom Duerig, Paul Natsev, Zoubin Ghahramani for his or her assist and assist.