Zero-Shot Robustification of Zero-Shot Models

Thank you for noting **the novelty of our method**, and the strength of our evaluations! * **On scenarios where language models fail**. There are two cases to examine. First, the reviewer asks about scenarios where the language model is not familiar with what causes, for example, CLIP, to fail. This is not a problem for our model: we do not need the LM to be aware of the intricacies and impacts of spurious correlations on a particular model. We only ask for relevant information about a task, not about a particular model's behavior on the task. Second, there are scenarios where the language model cannot provide any task-specific insights. In this case, our method cannot directly produce any improvements, because we do not have anything to use for determining what the harmful subspace should be. *However, even in this setting, all hope is not lost:* There are ways to **transfer insights** from one task to another. This requires a form of meta-prompting: we ask language models for what tasks are similar to the given tasks, and then ask for insights about all of these similar tasks. For example, suppose that the language model provided no insights whatsoever for waterbird and landbird categorization. We could simply ask the language model for what type of task (and task entities) are similar. We may receive an answer about distinguishing mammals from lizards as a potentially similar task. In this case, background would be a transferable insight, since mammal/lizard categorization is similarly not affected by background. There is a large space for such approaches.

Dear Reviewer, We thank you again for your feedback, questions, and suggestions! We believe we have answered all of your questions in our responses and the updated draft. If you have additional questions, we would love to answer them! The Authors

Thank you for noting the **novelty of our method**! We appreciate the kind comments and valuable feedback. We have make adjustments to figure captions as suggested by the reviewer in our updated draft. * **On the orthonormality assumption of concept embeddings**. We described these vectors as orthonormal purely for simplicity of explanation and derivation. However, **our method does not require the orthonormality assumption**. In general, if the $z_i$ are not orthonormal, we could describe them with a change of basis. The key requirement is that the subspaces where the helpful/harmful concepts live are not identical (i.e., as captured by the rank). Our method will fail if the subspaces of harmful concepts are, for example, identical to the subspace of helpful concepts. This is a weak assumption, however. Indeed, if it does not hold, harmful and helpful prediction concepts are totally aligned, so that good prediction is hopeless no matter what. We do not observe this to be the case in practice. In terms of the underlying theoretical model, we offer two arguments in support of our position. First, the empirical measurements (given below). Second, similar findings have been proposed in a variety of recent work [1, 2, 3, 4, 5]: [1] Trager, Matthew, et al. "Linear spaces of meanings: compositional structures in vision-language models." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023. [2] Chuang, Ching-Yao, et al. "Debiasing vision-language models via biased prompts." arXiv preprint arXiv:2302.00070 (2023). [3] Park, Kiho, Yo Joong Choe, and Victor Veitch. "The Linear Representation Hypothesis and the Geometry of Large Language Models." arXiv preprint arXiv:2311.03658 (2023). [4] Jiang, Yibo, Bryon Aragam, and Victor Veitch. "Uncovering Meanings of Embeddings via Partial Orthogonality." Thirty-seventh Conference on Neural Information Processing Systems. 2023. [5] Dev, Sunipa, et al. "OSCaR: Orthogonal Subspace Correction and Rectification of Biases in Word Embeddings." Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021. * **On Appendix F5 results**. The purpose of this experiment is to establish further evidence for our theoretical model. We sample what we expect to be *cleaner* or less contaminated datapoints expressing concepts (e.g., an image of all yellow pixels, an image of a square). In other words, we expect that these datapoints are likely to have smaller coefficients for irrelevant concepts (and fewer relevant concepts). We hypothesize that word embeddings with smaller harmful coefficient will be more similar in the embedding space with these *cleaner* images, and thus have a higher cosine similarity. To address the reviewer's concern on concepts simplicity, we perform the same experiment using CIFAR10 test set. |Class|Original embedding|Average embedding| |-|:-:|:-:| |Horse|0.258|**0.261**| |Automobile|0.238|**0.251**| |Dog|0.237|**0.240**| |Bird|0.248|**0.260**| |Cat|0.232|**0.245**| |Airplane|0.241|**0.262**| |Ship|0.244|**0.253**| |Truck|0.243|**0.248**| |Deer|0.243|**0.269**| |Frog|0.235|**0.256**| Indeed, even on more complicated concepts like those in the CIFAR10 labels above, the text embeddings with noise component averaged out have higher cosine similarities to the corresponding *cleaner* image embeddings. This result increases our confidence in the theoretical model. * **On the improvement claim.** Thank you for pointing this out! RoboShot produces average improvement of worst-group accuracy% of 15.95% with a trivial average decrease in average accuracy% of 1.44%. As suggested by the reviewer, have updated this number in the current abstract. We note that our findings are generally consistent with the robustness literature, where improving worst-group accuracy is the goal, and a decrease in accuracy overall is common. At the same time, we highlight that in several scenarios *both of thes metrices are improved.* We describe the reason for this in more depth below.

Thank your for pointing out the novelty of our method and the strength of our empirical results! * **On $X_{proj}$'s definition.** $X_{proj}$ is the sample embeddings post-application of the RoboShot procedure. We defined $X_{proj}$ in the parameters of Algorithm 2. * **On basis vectors ($z$).** The key strength of our method is that **it does not assume access to basis vectors $z$**. These vectors are part of our theoretical model only---but our approach does not require identifying them. In our model, we decompose embeddings as combinations of harmful, helpful, and benign components: $x = \sum_{s=1}^S \alpha_s^{\text{harmful}} z_s + \sum_{r=S+1}^{S+R} \alpha_r^{\text{helpful}} z_r + \sum_{b=S+R+1}^{S+R+B} \alpha_b^{\text{benign}} z_b.$ The goal of our procedure is to reduce $\alpha_s^{\text{harmful}}$ and increase $\alpha_{r}^{\text{helpful}}$ (the coefficient of harmful and helpful basis vectors). Since indeed these cannot be directly identified (i.e., we never know what $z_1$ is, for example), we query LLMs in the hope of text whose embeddings produce some information about these. Specifically, we obtain text from the LLM that we embedded. Such embeddings ideally span some subspace of the harmful $z$s that can be used to decrease $\alpha_s^{\text{harmful}}$ (and similary for the helpful concepts.). This does not ever require us to identify what $z_1, z_2, \ldots$ are. * **On standard classification tasks.** As suggested by the reviewer, we performed an experiment on ImageNet (using CLIP ViT-B-32). The performance of our method is roughly the same (we saw a ~1.1% difference). We do not expect RoboShot to produce improvement in this setting---we anticipate that many of the CLIP models have seen ImageNet or variants at training time and that no substantial biases are present that can be corrected. Fortunately, the result indicates that RoboShot largely retains performance in standard classification settings. * **On the number of insight vectors needed**. To address the reviewer's question, we do an ablation experiment and provide a theoretical characterization on number of insights needed. We provide the deatils in this response, and updated our manuscript with Appendix F7 that provides *empirical ablation results for Waterbirds an CelebA datasets*. As expected, we **observe that performance improves as the number of insights increases**. First, we point out that individual insights have different qualities---further suggesting the need for multiple insight vectors. The folowing table shows the different performance results from using one individual insight on harmful vector rejection |Insight pair| AVG% | WG%| |-|:-:|:-:| ['a bird with coastal migration', 'a bird with inland migration']|76.0%|26.0% ['a bird with keratin feathers physiology', 'a bird with hydrophobic feathers physiology']|80.8%|29.3%| ['a bird that eats bugs.', 'a bird that eats mainly fish.']|79.6%|30.2%| ['a bird that lives in watery environments', 'a bird that lives on land.']|86.2%|52.2%| On the theoretical characterization: Increasing the number of insights is beneficial until the span of insight vectors becomes the subspace of helpful/harmful features. Thus, for optimality, we need insight vectors up to the rank of the helpful/harmful subspaces. A synthetic experiment where we vary the number of insights validates this hypothesis. Here, the number of helpful/harmful/benign concepts is 12 respectively, with the embedding dimension 36 (12 + 12 +12). We sequentially increase the number of insights based on a similar synthetic experiment setup in Appendix F.2. In Appendix F.7 Figure 8, We can observe that the AVG/WG performance improves until the number of insights is the same with the rank of harmful/helpful subspace. After this, due to the presence of noise in superfluous insights, helpful components are removed / harmful components are added, which results in a performance decrease. This suggests that *there is indeed a sweet spot for the number of insight vectors*. While our method works particularly well if each insight from the LLM exactly targets a helpful/harmful component, there can be multiple insights for each concept. To resolve this issue, our ongoing work includes the application of robust subspace learning, replacing the QR decomposition step. This step uses the entire subspace that insight vectors span. In constrast, robust subspace learning only use the subspace that has strong signals to mitigate the challenges induced by noise [1]. We are also developing a robust version of the theoretical characterization---but we leave this for future work. [1] Lerman, Gilad, and Tyler Maunu. "An overview of robust subspace recovery." Proceedings of the IEEE 106.8 (2018): 1380-1410.

Thank you for noting the **novelty of our work** and the strong evaluation! * **On large scale datasets like ImageNet**. Our technique *does not require training*---so *large datasets are not a problem*. In general, the costs of our method are quite low for any dataset size. We obtain the insights and embed them, which is performed just once per dataset. These embeddings are re-used for modifying each datapoint's representation, which comes at only a marginal cost on top of inference. Indeed, some of the datasets in our experiments are reasonably large (e.g., CivilComments-WILDS has 133782 samples, Amazon-WILDS has 90078 samples). We list all dataset details in Appendix D1. The reason we did not focus on ImageNet in particular is that we do not anticipate our method to produce any substantial improvements for this dataset. This is to be expected for two reasons: (i) CLIP-style models already perform well on ImageNet, and (ii) it is likely that ImageNet or variants are in the training data used for these models, so that biases or spurious correlations are not as harmful. We verified this by running our method on ImageNet, where we observed similar performance as with vanilla zero-shot classification. * **On zero-shot classification versus other tasks**. We anticipate that our technique is usable in such settings as well. In general, the goal of our technique is to remove some harmful aspects of pretrained models (and boost some helpful ones) --- this is done by the harmful feature rejection (Algorithm 1 line 3-4), and useful feature addition (Algorithm 1 line 7-8). The overall difficulty of the task is not the crucial ingredient, but rather the presence or absence of these components. For our ongoing work, we are investigating using our technique in concert with segmentation. * **On the zero-shot terminology**. We agree that there are a variety of uses of the term "zero-shot". The usage that we refer to is using a pretrained model **without** (1) modifying the prompt with additional examples (i.e., few-shot prompting), (2) fine-tuning a pretrained model with labeled data, (3) continual pretraining with additional domain-specific data. All of these methods are commonly used for task-specific predictions. This usage is consistent with much of the literature. For example, CLIP's authors describe CLIP in the following way: "We ... turn CLIP into a zero-shot classifier."

Dear Reviewer, We thank you again for your feedback, questions, and suggestions! We believe we have answered all of your questions in our responses and the updated draft. If you have additional questions, we would love to answer them! The Authors