mBLIP: Efficient Bootstrapping of Multilingual Vision-LLMs

Thank you for your review.

(if will be released)

Yes, all our training data along with all trained models (full weights) will be publicly released.

The major concern is the novelty. In short, the work can be summarized as (1) using machine translation to translate image-text pairs (in English) to different languages, and (2) replace the language decoder in BLIP2 with a multi-lingual one and train the model with translated dataset.

Please see our general response to all reviewers.

The mixture of task highlighted in the paper as a contribution to the success of the model, although not considered in BLIP2, was actually explored in PALI. This fact further deteriorates the novelty of this work.

It would appear that we forgot to cite Pali when motivating our task mix in §3.2; we will fix this. We did however cite InstructBLIP, another Vision-LLM trained with a task mix, so we did not intend to claim this as a particular novelty of our approach.

However, unlike Pali, who use a rigid (and English-only) input format for instructions, we use diverse templates and additionally incorporate translated user instruction data (from Llava) so that our mBLIP models can handle a wider range of user inputs as illustrated in Figure 3.

It's not a surprise that task mix helps most in VQA and VNLI tasks because such an objective in included in the pre-training stage. The mixture of task actually hurts the captioning performance.

When comparing the results in Table 3 (row 2 for captions only, last row for the task mix), we do not see that the task mix hurts captioning performance: while English results for XM3600 and xFlickrCo are slightly lower, multilingual results are higher.

Also, VNLI is not included in the task mix, which is why our model has poor zero-shot results there (in zero-shot, i.e., without fine-tuning for VNLI, it never predicts the neutral class).

The model can be trained on 4 cheap RTX-3090 GPUs, which is great. But this claim in the conclusion has nothing to do with the contribution of the work, i.e., any similar models will enjoy the speed-up with 8-bit quantization and LoRA PEFT.

First, while LoRA and 8-bit quantization do contribute to the training efficiency, the core driver of our efficient training is the (comparatively small) size of our training data and our task mix – precisely because we re-align a pre-trained “English” Q-former to a new, multilingual LLM, we are able to achieve multilingual performance competitive to that of much larger models that have been trained from scratch with orders of magnitude less training data.

Second, the fact that our approach works with other models should be seen as a positive: if next week, a new SOTA multilingual LLM is released, we have a proven computationally effective method of inducing a corresponding Vision-LLM.

The paper does not state the data portion used to pre-train the model. In the model pre-training, 3 tasks are considered, but the captioning dataset looks much bigger than VQAs. Meanwhile, the small size of VQA could contribute less to comprehensive pre-training, but only provide some demonstration to downstream QA tasks (that's probably why QA and VNLI tasks perform well).

We do report this. In §3.2, we refer to Appendix C.1 for details of our training data where table 5 lists the exact number of images and samples used from each dataset.

As can be seen there, we limit the caption datasets to prevent this very problem you describe where the larger caption dataset would (proportionally) overshadow the other tasks.

The paper claims that based on a few prior works, the re-alignment does not need larger dataset. However, some of the listed prior works (on arXiv) have very weak quantitative studies, which may not support authors' claim. Thus, such a claim has to be validated by ablation studies like re-alignment with different size of pre-training data, e.g., 3M, 12M, 120M, 400M, etc.

We cite two works to support this claim. We can agree that Zhu et al., 2023 lacks proper quantitative evaluation; Zhang et al., 2023, however, features extensive experiments that clearly indicate that one can get comparable results with re-alignment on far less data.

While the suggested experiment would be interesting – we do not dispute that with more data, you can (very likely) get better results, but we also suspect that the gains would diminish logarithmically with more data. The proposed data-scaling experiment/ablation is also too expensive for us to run: training with 100M examples would take at least 40 days on our compute.

In addition, there is another problem: for the greater data scale (>100M images), only image-caption data is available, which would greatly skew the proportions in our task mix.

Thank you for your review.

The paper does not seem to have a strong technical contribution. On the pro side, I like the application, and there's some analysis of the impact of the different pieces in Table 3, which offsets somewhat the lack of technical contribution.

Please see our general response to reviewers.

I am a bit unsure about the impact of LoRA. A task prompt has much lower capacity to adapt to any downstream task compared to adding LoRA (especially if LoRA adapters are added to every 1x1 layer). So I'm wondering if part of the performance is due to LoRA vs no LoRA.

We are not sure what you refer to with “task prompts”? Do you mean (1) InstructBLIP’s method of including the prompt in the Q-former for task-contextualized visual tokens or (2) prefix-tuning, i.e., the trainable prompt-token embeddings as proposed in [1,2]?

Considering the latter more likely – in the LoRA paper, the authors demonstrate that LoRA significantly outperforms prefix embeddings, as well as the other parameter-efficient fine-tuning (PEFT) approaches, such as bottleneck adapters or BitFit. They also show that it is the only PEFT method that consistently matches (or surpasses) the performance of full fine-tuning. For all these reasons, LoRA is arguably the most widely used PEFT technique in NLP. LoRA fine-tuning of the multilingual LLM in our approach indeed contributes to the overall performance, as demonstrated by the ablation from Table 3 (row 3 vs. last row): LoRA fine-tuning of the multilingua LLM yields better results than when the LLM is kept frozen.

Also, if there is enough data to train LoRA adapters, why then there is no data to train a task prompt?

We are somewhat puzzled by this question and are not sure what exactly you refer to. We are not aware of making a claim that there is not enough data to train a task prompt anywhere in the paper.

If by task prompt you mean InstructBLIP’s method with prompt-contextualized visual tokens, then the problem is the one described in §3.1: the Q-former is BERT-based (i.e., a monolingual English Transformer Encoder), not capable of handling multilingual prompts (i.e., prompts in languages other than English).

If (which we believe is more likely) by task prompt you refer to prefix embeddings (prefix-tuning) in the vein of [1,2]: then there is no conceptual obstacle to doing prefix-tuning instead of LoRA, but it would (1) lead to inferior performance to LoRA (as stated above, LoRA has been documented to outperform competing PEFT methods and is de facto the default PEFT approach at the moment in NLP), and (2) increasing both training and inference time, due to the increase in sequence length (from additional prefix embeddings) and quadratic nature of the attention mechanism in the Transformer. So while prefix-tuning is conceptually viable PEFT technique for our setting (but so are the bottleneck adapters or BitFit), there is no reason to use it over LoRA. We are happy to emphasize this in the manuscript (advantages of LoRA over competing parameter-efficient fine-tuning methods)

is "re-aligning" actually fine-tuning?

In a way, yes: The (multilingual) LLM is fine-tuned in a parameter-efficient manner with LoRA; the Q-former is first aligned to the LLM (similar to stage two in BLIP-2) and then further fine-tuned with the task mix (similar to the InstructBLIP fine-tuning of a BLIP-2 model) in one training step. However, to better illustrate that our approach re-aligns a Q-former from one LLM to a new (multilingual) LLM and to separate this step from the classic task-specific fine-tuning in our downstream evaluation (for xGQA, XVNLI, & MaRVL) we decided to use ‘re-align’ instead of ‘fine-tune’ for the training on our task mix.

can the use of LoRA be better justified?

As we motivate in §3.1 and show in the ablation, training the LLM yields better results than keeping it frozen. Full fine-tuning of all parameters of the LLM is too computationally expensive and one must resort to parameter-efficient fine-tuning (PEFT). LoRA is, as discussed above, widely accepted in NLP as the best PEFT technique – outperforming the competing approaches (bottleneck adapters, bitfit, prefix-tuning) and often matching the performance of full fine-tuning.

[1] Li & Liang, 2021. Prefix-Tuning: Optimizing Continuous Prompts for Generation

[2] Lester et al., 2021. The Power of Scale for Parameter-Efficient Prompt Tuning

Thank you for your review.

There are still big differences of accuracy between English and other languages on XM3600 and xFlickerCo testing.

For xFlickrCo, we are comparing against English-only monolingual models based on Flan-T5. Monolingual language-specific LLMs (especially if the language is English, the most resourced language there is) generally outperform multilingual counterparts for that particular language. This phenomenon is known as the curse of multilinguality [1,2]. English-only models are in general better in English than multilingual models so the gap in CIDEr score is to be expected there.

For XM3600, the English CIDEr scores seem to us qualitatively similar to the multilingual results: our mBLIP variants outperform Pali-X (zeroshot), LMCap, and Thapliyal et al. (2022) and are outperformed by the Pali models that have been fine-tuned on MSCOCO.

In addition, Table 2 is not so clean to understand.

We are aware that Table 2 contains a lot of information but we tried our best to explain everything clearly in the caption (and the text of discussions associated with the Table). Is there anything in particular that was hard to understand so we adjust the table accordingly?

[1] Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., ... & Stoyanov, V. (2020, July). Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 8440-8451).

[2] Pfeiffer, J., Goyal, N., Lin, X., Li, X., Cross, J., Riedel, S., & Artetxe, M. (2022, July). Lifting the Curse of Multilinguality by Pre-training Modular Transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 3479-3495).

Thank you for your review.

One weakness is to name a translated dataset as a multilingual dataset, which doesn't cover the real benefits of multilingual at all (i.e. multiple cultures, examples from different locations from local people speaking those languages). This is misleading as the future work will use the translated dataset which will reinforce the emphasis of English centric research, under the translated "multilingual" directions.

We fully agree with the reviewer, but want to point out that this is, unfortunately, a general problem in multilingual vision-language models. For example, CCLM, Ernie-UniX2, or UC2 all also train with machine-translated data. PaLI models are trained with both machine-translated data and (proprietary) multilingual web-crawled data. There is multilingual caption data available in Laion5B but these are noisy web-crawled captions. For high-quality captions and other tasks like VQA, there is only English data available.

However, our evaluation setup does take this problem into account and includes test datasets that require both multilingual and multicultural knowledge: XM3600 in particular selected and annotated images from 36 countries with speakers from those countries, and MaRVL is built from the ground up (and also annotated by native people) to include images and concepts from 5 different countries.

Nonetheless, you make a valid point and we will make sure to add a short discussion on true multilingual data vs. “multilingual” data obtained through translation and disadvantages/pitfalls of the latter.

The idea of applying Q-Former to LLMs is not novel and has been explored by the previous BLIP papers. I have not found BLIP baselines reported (at least on EN) in the paper (except for the right sub-table of Table 1). More comparisons with BLIP baselines, especially on the impact of multilingual data to the original English tasks would be nice to show the impact of this paper.

The comparison with BLIP-2 models is difficult because their underlying English LLMs are better on English tasks than multilingual models. Monolingual language-specific LLMs (especially if the language is English) generally outperform multilingual counterparts for that particular language. This phenomenon (i.e., the downside of sharing representation space across many languages) is known as the curse of multilinguality [1,2].

We did consider including the zero-shot GQA results for BLIP-2 and InstructBLIP Flan-T5-XL results in Table 2 (44.0 and 48.4% accuracy, our mBLIP mT0-XL has 42.6) but decided against it for the sake of readability of Table 2, which is already quite full with information.

The comparison of mBLIP with PaLI is quite unclear. Is mBLIP a zero-shot model or finetuned model? If it's the latter then it doesn't make any sense to claim that mBLIP outperforms (zero-shot) PaLI.

This is a good question with no straightforward answer:

PaLI-X is pre-trained with (amongst other data) millions of captions translated to the 36 XM3600 languages. For fine-tuning, they are then trained on ~5M examples of MSCOCO again translated to the 36 languages.

mBLIP is trained with our task mix for 2 epochs, which includes 400k MSCOCO captions automatically translated to 95 languages.

So on one hand, mBLIP has seen data for some of the downstream fine-tuning tasks in its re-alignment training (but orders of magnitude less than the fine-tuned PaLI) but on the other hand, we did not optimize our setup for the 36 XM3600 languages (and PaLI does).

To make the comparison more complicated, our mBLIP models have 5 to 8B parameters of which we train only 128M. Pali-X has 55B parameters, all of which are updated during fine-tuning training.

In sum, this means that mBLIP has some advantage regarding the (type of the) training data, whereas Pali-X has an advantage in terms of training languages (matching exactly those from XM3600), model size, and size of the (pre)training data. We thus believe that it is overall fair to compare mBLIP against Pali-X (zero-shot).

References:

[1] Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., ... & Stoyanov, V. (2020, July). Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 8440-8451).

[2] Pfeiffer, J., Goyal, N., Lin, X., Li, X., Cross, J., Riedel, S., & Artetxe, M. (2022, July). Lifting the Curse of Multilinguality by Pre-training Modular Transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 3479-3495).