Prompting Language-Informed Distribution for Compositional Zero-Shot Learning

Thanks for reviewing the paper. Below we present our responses to the raised concerns.

• Rely on the LLM quality. We would clarify that leveraging LLMs for the CZSL task is one of our proposed ideas in this paper, and it has never been explored in the CZSL literature. Therefore, whether the CZSL performance replies on LLM quality or not is an open research question, and interestingly, we found the performance is not always better if a better LLM is utilized (see Table 3). We believe in the future, there will be more research investigating how to better utilize LLMs to solve the challenging CZSL task.
• Impact of VLPD. The VLPD primarily improves the open-world CZSL performance ($H_{ow}$ and $\text{AUC}_{ow}$), rather than the closed-world ones. This can be observed in Table 2 which indicates significant open-world performance gains. This observation is expected because the VLPD is developed to mitigate the overfitting issue on the visual data of seen compositions. Thus, in a much larger open-world decision space where most compositional class names might be infeasible, a less overfitting model w.r.t. the seen compositions is expected to achieve higher open-world performance.
• Inter-class correlation. We clarify that the proposed LID is simultaneously minimizing the intra-class variance and maximizing the inter-class separability as stated in the Lines below Eq. (1) (the last 3 lines of Page 5). More specifically, the learning goal is to minimize the NLL upper-bound (r.h.s of Eq. (1)). This will encourage minimizing the $h_{k,y}$ term which is determined by the positive intra-class variance terms $\Sigma_{kk}$ and $\Sigma_{yy}$, and also the negative inter-class terms $-\Sigma_{ky}$ and $-\Sigma_{yk}$. It's thus clear that inter-class correlation is also optimized.

We sincerely thank the reviewer's acceptance of this paper and all the efforts in reviewing it.

1. $N$-view augmentation

We would clarify that the comparison is fair though it is not used in comparison methods. The reason is that the N-view augmentations on the images correspond to the motivations of our distributional modeling, that introducing variety/uncertainty into visual-text alignment by TFE and VFE leads to better zero-shot generalization. This is the contribution of our method that the VFE depends on, rather than a general technique that can be adopted by other methods.

To further validate its effectiveness, we compare the model without $N$-view augmentation with the SoTA method DFSP as shown below. It shows that our method still performs better than (or comparable with) DFSP.

2. Performance of LID

Note that the LLM is the core component to form the LID. Besides, using LLM-generated text as class distribution points in prompt learning is a new paradigm for zero-shot learning, which has never been explored in literature, rather than incremental engineering. The performances of LID and LLM are individually analyzed in the response to the Reviewer BPd3 (see the Table in response), which shows the contribution from both especially in the closed-world CZSL setting.

3. Figure 2

Note that the figure has provided the LID part (top-center part). We will revise it to make the LID more noticeable for readers.

Thank you again for your constructive comments!

I thank the authors for their response. I would like to ask for further clarifications:

1. From the table, the MLP (x1) row achieves better results than the full model in 3 out of 4 metrics. Is it using SLM and VLPD? and what happens if the single MLP is replaced by the single cross-attention layer (without SLM and VLPD)? I am asking this question because the contribution (b) at the end of the introduction depends on the impact of these two components.
2. The ablation is reported on MIT-states but a) the results of all models are very close and b) this dataset is noisy (Atzmon et al. 2020). As I mentioned in point 1, it would have been helpful/made the contribution stronger to verify if the improvement brought by each component holds across datasets as well. Is there a reason for performing the ablation only on MIT-states?
3. While I appreciate the will to update the manuscript to improve its clarity, it would have been better to update the article as well (given that ICLR allows the upload of revisions). Is this due to time constraints and/or some changes that are hard to incorporate?
4. It would also be helpful answering to the concerns of Reviewer 4PfY, pointing to potential mismatches in the number of views in the comparisons.

1. The MLP (x1) row keeps all others the same (uses SLM and VLPD) except for the TVE/VFE module, where we replace the cross-attention with MLP. Though it performs better, the improvement is rather small (less than 0.1 points). This is why we conclude that the type of the structure does not matter. For such structural exploration on the model without SLM and VLPD, we can do that but the inner structure of the module is NOT our claimed contribution and it would take more time at this moment. For your concern about our claimed contribution (b), we already have the variable-controlled experimental results in Table 2 (row (f)(g)) that validate the VLPD and SLM.

2. In terms of the ablation study on more datasets, we are currently doing the ablation study on the UT-Zappos dataset. At this moment, we already have the new results by individually removing LLM, LID, VLPD, and SLM which are the major claimed technical contributions.

This table shows that the proposed components indeed work well on the UT-Zappos dataset. We will add this Table to the main paper.

3. Yes, we appreciate your understanding of the time constraints and workload needed to improve the clarity. Actually, we just finished the CVPR'24 submission work and put all my effort into the ICLR rebuttal and experiments these days. We promise that the substantial revision to improve the clarity will be done later.

4. For the reviewer 4PfY's question, we clarify that there is no such mismatch problem. The reason is that the N-view augmentations on the images correspond to the motivations of our distributional modeling, that introducing variety/uncertainty into visual-text alignment by TFE and VFE leads to better zero-shot generalization. This is the contribution of our method that the VFE depends on, rather than a general technique that can be adopted by other methods.

We sincerely thank the reviewer's support for the principles and ideas behind the approach. The comments are constructive to improve the quality of this work.

1. Experimental validation of design choices.

We would like to emphasize that the design choices of TFE, VFE, $f_s$, and $f_o$ are not the contribution of this paper. To address the concerns about their impact, here we provide the results of the model with different design choices in the Table below. Specifically, for simplicity, we keep the structures of TFE and VFE the same and replace their default one-layer cross-attention (XAttn$\times 1$) implementations with three-layer cross-attention (XAttn$\times 3$) or one-layer MLP (MLP$\times 1$). For the SLM module, we replace it with the version that only uses compositional logits (comp.-only logit), and the version with simple weighted averaging (w-avg. logit) by setting $\lambda$ to 0.1 (the same as the expectation of the Beta distribution in our SLM).

We have the following observations.

• LLM indeed has a noticeably positive effect while the cost is acceptable. For the CZSL evaluation especially the very challenging open-world metrics, it is normal that the scales of numbers are very low and even 0.5 points are meaningful in practice (see other methods in Table 1). If removing LLM descriptions (w/o LLM, w/o LID), we present results in Table above (row (1)) which shows the closed-world performance decreases significantly ($-1.12$ points of $AUC_{cw}$). With LLM added but still without LID, the closed-world performance further improved ($+0.67$ points of $AUC_{cw}$). These results demonstrate that LLM and LID could collectively improve the CZSL performance. Regarding the concern on the cost of LLMs, in practice, the compositional descriptions are generated offline by frozen LLM and their features are pre-extracted by frozen CLIP so that the cost for our model training and testing is acceptable.
• Both the structure of TFE/VFE and the introduced distribution matter. Comparing the result row (2)(3)(4) with our full model (the last row), it shows that without the distribution modeling by LID (LLM, w/o. LID), the results (row (2)) are worse than models with LID but instead using simple MLP (row (3)), while better than the model with more cross-attention (XAttn$\times 3$) layers (row (4)). This indicates that the number of TFE/VFE layers matters more but the types (attn/mlp) matter less than the LID. Since LID has never been explored in CZSL literature, these findings could inspire future exploration on how to better utilize LLM-based distributional semantics.
• SLM is better than simpler (or without) logit aggregation. In row (5), we simply use compositional logits without primitive ones, while in row (6), we use simple weighted averaging. We see that neither of these simple methods could achieve better performance and noticeably, only using compositional logits lags far behind our SLM strategy which indicates the importance of primitive-based logits in the CZSL task.

2. Presentation and Notations

Thanks for providing the valuable writing suggestions! We will revise the Abstract and Introduction text part to convey clearly the methodological idea. For the notations, we have intended to precisely convey the technical truth of our method through the notations, while not being aware of their overloaded information for readers. Thanks to your suggestion, we will try our best to simplify them. We will also revise the figures in the final version.

We thank the reviewer for viewing this paper.
In terms of the novelty of our motivations and methodology, we would like to emphasize the key factors that might be ignored by the reviewer:

• On the general ZSL problem, we summarize the two aspects, i.e., context diversity addressed by modeling class distribution over soft prompts and the informativeness inherited from LLM-based language prompts, are only independently studied in the literature. In this paper, we advocate that they can be naturally bridged together to take the merits of both while overcoming their limitations. This points to a brand-new prompt learning paradigm, that is general to any vision-language foundation models on zero-shot classification problems.
• Specific to the CZSL task, we are the first to model the class-wise distribution over both the compositional and primitive classes. Besides, our VLPD is different from existing CZSL literature as we decompose both image and text modality rather than simply decomposing text features in the DFSP and HPL (see text below Eq.(2) and Table 5).

For the technical questions:

• Our augmentation to image input is NOT a test-time adaptation, as its relevant module VFE has to be learned in training.
• For the Eq. (1), the matrix $A$ is defined to represent pair-wise class correlation on each of $d$ features so that its shape is $d\times C\times C$ where $C$ is the number of classes. Therefore, the $A_{k,y}$ is a $d$-dimensional vector. Note in practice, before the marginal logit $v^\top A_{k,y} {v}$ is computed, $A_{k,y}$ is diagonalized to have the shape $d\times d$. To avoid misunderstanding, we will update the marginal logit as $v^\top `diag`(A_{k,y}) v$ in the main paper.
• As requested, results w.r.t. the value of $N=0$ are shown in the Table below. It clearly shows the effectiveness of the VFE that takes an image with $N$ augmented views as input.
• The final training loss is a weighted sum of $\mathcal{L}_y$ in Eq. (1) and $\mathcal{L}_s+\mathcal{L}_o$ in Eq. (2). The weights between them for different datasets are summarized in Table 7 of the supplement.