NeurIPS 2020

Prophet Attention: Predicting Attention with Future Attention


Review 1

Summary and Contributions: From the motivation that existing methods on image captioning methods have a "deviated focus" problem, which attention-based decoder calculate the attention weights used on previous words instead of the one to be generated, this submission proposes a method to regularize such attention weight using future attention weights

Strengths: - Easy to follow writing. - Thorough experiments with ablation study - SOTA performance

Weaknesses: - While I enjoyed the introduction and motivation, the implementation of solving "deviated focus" problem seems straight-forward (i.e., average of future attention weights, POS tag words and split the cases) - It would be nice to have ablation on loss function for attention regularization using L2 norm, KL divergence, etc instead of L1 norm. Why did you choose L1 norm? ============================================================ Author feedback clarifies my questions.

Correctness: It seems correct

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: No

Additional Feedback:


Review 2

Summary and Contributions: This paper focuses on the "deviated focus" problem on attention-based image captioning models. Specifically, the attention weights are calculated based on previous words instead of the one to be generated. Thus, the authors calculate attention weights from the "future words" to regularize the "deviated" attention. Extensive results on Flickr30k Entities and MSCOCO datasets have demonstrated the effectiveness of the Prophet Attention.

Strengths: 1. The whole paper is well written and easy to follow. 2. The proposed Prophet Attention mechanism is model-agnostic and can be easily incorporated into different image captioning models and other tasks (ie, paraphrase and video captioning). 3. The model achieves a new state-of-the-art performance on two benchmarks.

Weaknesses: ========= Post-rebuttal Edit ============= After reading the authors' response and all other reviewers' comments, I choose to keep my rating as 6 (Weak Accept) for the following reasonings: 1) The proposed paper is well-written and all strong results have demonstrated the effectiveness. 2) The whole model is simple, straightforward, and experiments-oriented. More comprehensive considerations about different situations or words (like the mentioned verb) can be better. ==================================== 1. The ideas are partially not convincing, the attention for some words should not only from the future, but also from the history. As mentioned in Eq. (9), the Dynamic Propohet Attention considers three types of words: noun phrase, non-visual word, and others. For example in Figure 1, when generate some verb "holding", based on the current formulation, they belong to the "others" category. And the \hat{\alpha} (cf. Eq(5)) come from the the next future word. However, intuitively, it is more reasonable to focus on both the subject and object of "holding", \ie, "a woman" and "a yellow umbrella". Analogously, for "wearing", the targeted attention should focus on "a woman" and "a yellow coat".

Correctness: Yes. Although I think the attention of some words should focus on both history and future (see Weaknesses part), in many cases, the attention should be consistent with the current generated words.

Clarity: Yes. This paper is well written and all figures are well-illustrated.

Relation to Prior Work: Yes. This paper have a clearly discussion about the difference between existing attention-based models and attention supervision on other vision-and-language tasks.

Reproducibility: Yes

Additional Feedback: 1. For the first visualization results in Figure 3, the model with DPA can predict smiling with the purple box, is it should be attend to the whole boy (ie, "a smiling boy")?


Review 3

Summary and Contributions: Different from previous attention-based models usually use the hidden state of current input to attend to the image regions, this paper proposes a Prophet Attention to enable attention models to correctly ground words to be generated to proper image regions. The proposed model achieves great success on image captioning. -----------update after rebuttal -------------------- Authors have addressed all my concerns in the rebuttal and I vote for acceptance of this submission.

Strengths: The motivation for this paper is clear and soundness. The authors identified a significant problem of current attention-based models (for image captioning) and propose a simple 'trick' to address it. The proposed method is easy to follow and implement. And several experiments have been done to prove its effectiveness. The paper is clear and easy to follow.

Weaknesses: They are not weaknesses but I have some concerns: 1. I do agree with the authors that the 'attention' should not be calculated based on the previous states, instead, it should use the 'future' states. However, according to the setting of hyper-parameter 'lambda', the model only works well when it equals to 0.1, which is a rather small value. This may suggest that this 'modification' of the attention mechanism is not significant. I can accept the explanation that a large 'lambda' will lead to the attention weight bias towards the last token in the sequence, but whether this suggests the 'nested' regularization mechanism proposed in this paper is not appropriate or even a wrong direction? 2. It is OK to say the proposed model outperforms all the published state-of-the-art models, however, I don't think it is appropriate to say it achieves the 1st place on the leaderboard. It only achieves the best results when C40 references are used. And it is not clear whether other models use ensemble or not. 3. This will be an excellent work if authors can show it also works on other attention-based models, on non-caption generation tasks.

Correctness: The method is simple but looks effective.

Clarity: yes, very clear and easy to follow.

Relation to Prior Work: yes, the differences are clear.

Reproducibility: Yes

Additional Feedback:


Review 4

Summary and Contributions: In this paper, the author focus on correctly grounding the image regions with generated words in the attention model, without any grounding annotations, additional parameters and extra inference computations. Importantly, they propose the Prophet Attention, which is similar to the form of self-supervision for calculating attentional weights based on future information, and force the attention model to learn to correctly ground each output word to proper image regions. In my view, the contributions are: 1. a new attention mechanism to correctly ground words to proper image regions. 2. achieves a new good performance. 3. this strategy is general and can be applied to other similar areas. In particular, I am glad that this work achieves a new 1st place on the COCO leaderboard.

Strengths: 1. As written above, this method achieves a good performance. 2. Figure 2 is pretty clear for understanding. 3. It is robust to perform experiment on baselines: Up-Down, GVD and AoANet. 4. Experiment evaluation are adequate and compact. 5. This method can be applied to other similar areas.

Weaknesses: 1. The paper is experimental oriented, and there is not enough theoretical analysis and grounding. 2. The paper does not explain the influence of introducing structure on time and lacks relevant analysis. How much the model slow down? 3. The idea is not novel, since incorporating future information has been studied and this paper has not presented. 4. There are many grammatical and clerical errors in the text.

Correctness: Pretty correct.

Clarity: not up to much, there are a lot of mistakes, e.g. line 220 Table 4 should be Table 6.

Relation to Prior Work: Incorporating future information is not new in image captioning task, for example, [1] [2][3], however, the author has not compared. I think author should illustrate the difference. [1] Modeling Future Cost for Neural Machine Translation [2] Qin, Yu, et al. "Look back and predict forward in image captioning." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019. [3] Ren, Zhou, et al. "Deep reinforcement learning-based image captioning with embedding reward." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.

Reproducibility: Yes

Additional Feedback: This is a nice article and will make a good contribution to the image captioning community. I hope the author can select several weakness as feedback.