LEMoN: Label Error Detection using Multimodal Neighbors

Thank you for the insightful review and constructive feedback!

I noticed that the most recent dataset used was published in 2020. Is the method still effective on datasets published in the last two years? Do mislabeled data still exist in these recent datasets?

We clarify that we evaluated our method on CC3M [1] (from 2021) in Appendix I.9 and the DataComp benchmark [2] (from 2023) in Appendix I.10. As we have motivated in the introduction, and also has been highlighted in many prior works [2-4], the issue of mislabeled data is only growing with time due to the use of billion-sample scale datasets collected from scraping the web.

Finally, we note that the published work which the reviewer later references [5] uses no datasets from later than 2015.

The author compares the method with LLaVA but does not provide fine-tuned results for LLaVA. Since it is known that mislabeled data degrade model performance, and LLaVA does not address mislabeled data through fine-tuning, the comparison with large models lacks persuasiveness.

The study does not provide results on fine-tuning large models such as LLaVA. Since mislabeled data can degrade model performance, it is important to examine whether filtering with LEMON improves the performance of fine-tuned large models. A comparison with fine-tuned LLaVA would strengthen the persuasiveness of the results.

We have already conducted several experiments showing that filtering with LEMoN improves the performance of downstream large models. In particular, we have finetuned GenerativeImage2Text models in Section 6.2, pretrained CLIP models on MIMIC-CXR in Section 6.4, pretrained CLIP models on CC3M in Appendix I.9, and pretrained CLIP models on DataComp in Appendix I.10. We believe this sufficiently addresses the reviewer's concern.

The author only considers image and text modalities. However, could the method generalize to more widely used modalities such as video and audio? Can relevant experiments be provided to support this?

Additionally, the current work focuses on image and text modalities. Expanding the study to include other widely used modalities, such as video and audio, would help demonstrate the generalizability of the method. Conducting experiments on multimodal datasets beyond image-caption pairs could further validate LEMON’s effectiveness in broader applications

We believe that demonstrating LEMoN's effectiveness on image-text pairs is a sufficient contribution, and extending it to other modalities is out of scope for this paper. We will note this as an area of future work in the revision.

While the paper claims novelty in using multimodal scoring for label noise detection, a similar approach has recently been explored in "VDC: Versatile Data Cleanser based on Visual-Linguistic Inconsistency by Multimodal Large Language Models," ICLR 2024.

We have already compared our method to this baseline in Table 2. We strongly dispute the claim that VDC is a "similar approach" to LEMoN. VDC entirely relies on prompting LLMs and VLLMs. In contrast, our method does not utilize any prompt engineering, and instead utilizes the neighborhood information in image and text representations of contrastively pretrained models. As a result, not only does VDC perform worse empirically (Table 2), it also has much higher runtime (Table I.11). We emphasize that VDC does not utilize multimodal neighbors, or even embeddings at all, in any form.

[1] Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts. CVPR 2021.

[2] DataComp: In search of the next generation of multimodal datasets. NeurIPS 2023.

[3] Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv:2110.01963.

[4] What's In My Big Data? ICLR 2024.

[5] VDC: Versatile Data Cleanser based on Visual-Linguistic Inconsistency by Multimodal Large Language Models. ICLR 2024.

Thank you for the insightful review and constructive feedback!

In appendix A.2, authors only present the visualization results for classification tasks, which is a much easier case than tasks involving image captioning, where the text can be natural language and have more diverse space on J(x).

For assumption 2, though appendix A.2 shows the distribution of the whole dataset, it is much better to show the sample-wise visualization results, since assumption 2 is a sample-wise claim.

Thank you for pointing this out! To address these two concerns simultaneously, we conduct an experiment on captions from the flickr30k dataset. We select 20 random captions, then use Llama 3.1-8B-instruct to generate 50 paraphrasings of each caption (via sampling with temperature), corresponding to 50 samples from $\mathcal{H}(Y)$ for each caption. For the samples $Y' \not\in \mathcal{H}(Y)$, we randomly select 50 other captions from the dataset. To match the support of the Gaussian, we take the distance function to be the log cosine distance (note that this does not change the ordering of the score across samples). We compute this distance using the text encoder from OpenAI CLIP ViT-B/32, and plot histograms for each caption. The results are shown here. Running the same Shapiro-Wilk test from Appendix A.2, we find that of the positive samples, 8/20 are Gaussian, and of the negative samples, 16/20 are Gaussian. Thus, there is some evidence the Gaussianity assumption holds for natural language and complex paraphrase functions.

the conclusion on "our proposed multimodal neighborhood score, provides a better than random signal at detecting mislabeled samples" is tricky. As long as there are slight distribution differences over two distributions (say, with different $\mu$) in appendix A.2, this conclusion still holds, but does not provide significant insight into how the detection method is good enough.

We emphasize that Lemma 4.2 is a specialization of Theorem 4.1 meant to demonstrate that only loose conditions are necessary to obtain non-random signal. Our Theorem 4.1 provides the exact expression for the AUROC of the score as a function of the distribution parameters.

Could authors please show the experiment results on directly setting γ1=γ2=0 in the algorithm?

We have provided results for setting $\tau_1 = \tau_2 = 0$ in Table I.9.

However, it seems that the impact of noise level is not presented in the theoretical part, and authors directly claim that the proposed method can achieve better noise detection performance than the random signal case.

The impact of noise level in the neighbors is accounted for in Theorem 4.1 via the $p$ term.

Notice the experiment in Figure I.1 in Appendix for different noise levels, please explain the phenomenon and tendency on F1/AUROC on mother datasets.

To better match the theoretical setting, we examine the performance of individual scores $s_n$ and $s_m$, without the $d_{mm}$ term and with $\tau_1 = \tau_2 = 0$. Looking at the influence of $p$ in Theorem 4.1, we find that as $p \rightarrow 1$, the AUROC approaches 0.5. As $p \rightarrow 0$, taking distribution parameters to be fixed (i.e. $\mu_1, \sigma_1$, etc), the AUROC approaches a fixed constant, which, under the assumptions of Lemma 4.2, is greater than 0.5.

For $p \in (0, 1)$, the function is strictly decreasing in $p$. Thus, from the theory, we would expect the AUROC to be strictly decreasing with higher noise rate, going down to 0.5 for $p = 1$. Empirically, we do observe the decrease in AUROC, with a faster decrease for mscoco than mmimdb (which according to the theory is due to dataset specific parameters like $\mu$'s, $\sigma$'s, and the moments of $\zeta$).

Regarding variance: we would like to note that the result of our Theorem 4.1 is for the "population" AUROC without finite sample considerations. In practice, as in our experiments, AUROC is estimated using finite samples from a fixed dataset. The variance that is observed is due to the variance of this statistical estimator. The variance of empirical AUROC is related to the variance of a Mann-Whitney U statistic, and has been characterized in [1]. This statistical variance is independent of our theorem.

Finally, our theory does not provide an explanation for the F1 score. This is tricker to characterize theoretically, as F1 is computed given a particular threshold on the score. This threshold is selected to be the one that maximizes the F1, and the F1 is a non-concave function of this threshold.

Essential References Not Discussed

Thank you for pointing these out. We have added [2] and [3] respectively to address these.

[1] Confidence Intervals for the Area under the ROC Curve. NeurIPS 2004.

[2] Deep k-NN for Noisy Labels. ICML 2020.

[3] Detecting Corrupted Labels Without Training a Model to Predict. ICML 2022.

Thank you for the insightful review and constructive feedback!

Empirical performance on realistic noise. Downstream performance on CC3M and Datacomp is almost the same as Clip Similarity baseline.

First, we would like to emphasize that we evaluate our method against the baselines on several other datasets with human noise, including CIFAR-10N, CIFAR-100N [1], and StanfordCars and MiniImageNet [2]. These datasets contain noise from human annotations collected on Amazon Mechanical Turk and the Google Cloud Data Labeling Service respectively. Second, we would like to note that, in addition to CC3M and Datacomp, we also observe increased downstream performance of training on LEMoN filtered datasets in CIFAR-10N and CIFAR-100N (Figure 3) and mscoco (Table 3).

Empirical performance on realistic downstream tasks is unsatisfactory. Perhaps fine-tuning under limited data (few-show training samples) is better suited to evaluate these methods?

Thank you for this suggestion! We have conducted additional experiments for these CC3M-pretrained models by linear probing on the VTAB benchmark, first in the few-shot setting where we select 5 random samples per class, and next where we finetune on the standard training split of each dataset. Our results can be found here. Overall, we find similar trends as the zero-shot setting, where LEMoN marginally outperforms the baseline, with both underperforming the model that has been pretrained on the whole corpus.

Simpler ways to incorporate k-NN information. For ex. average CLIP score between the caption $y$ and neighboring images of $x$ (and vice versa). This avoids the $\tau_2$ hyperparam, but ignores paired information of neighbors.

As described in weaknesses, a discussion on simpler ways to incorporate k-NN information would be informative.

Thank you for this suggestion! We have added this alternate way of integrating neighbor information as suggested by the reviewer:

$s(x, y) = d_{mm}(x, y) + \beta s_n (x, y, \mathcal{D}) + \gamma s_m(x, y, \mathcal{D})$

with $s_n(x, y, \mathcal{D}) = \frac{1}{k} \displaystyle\sum_{j=1}^k d_{mm}(x_{n_j}, y)e^{-\tau_{1, n} d_{\mathcal{X}}(x, x_{n_j})}$

and $s_m(x, y, \mathcal{D}) = \frac{1}{k} \displaystyle\sum_{j=1}^k d_{mm} (x, y_{m_j})e^{-\tau_{1, m}d_{\mathcal{Y}} (y, y_{m_j}) }$

Which drops the $\tau_2$ hyperparameter as the reviewer suggested. To maintain fairness, we use the same model selection strategy and hyperparameter grid for remaining hyperparameters as $\text{LEMoN}\_{\text{opt}}$. We evaluate this alternate neighborhood score versus LEMoN, and these results can be found here. We find that LEMoN outperforms this alternate neighbor method on the majority of datasets.

Finally, we note that we have also discussed some alternate ways of integrating neighbor information in Appendix B when we compare LEMoN conceptually and empirically with a baseline which uses semantic neighborhood information for a different purpose. We will add more exposition to this discussion in the revision.

It is not clear if filtering noisy data is preferred to techniques that learn in the presence of label noise. Empirical comparison with [2] on captioning datasets would strengthen the paper.

When are filtration based approaches preferred over techniques that learn with label noise?

In our view, noise-robust training algorithms are a disjoint field of work from noisy label identification. In particular, identifying noise labels is a more flexible approach, with applications beyond just removing these samples for downstream model training. By identifying mislabeled samples, we can also characterize systematic errors or biases in datasets (such as in Figures I.4 and I.5), which can then be fixed, both by repairing existing data and improving future data collection practices. This is especially important for practitioners looking to release high-fidelity datasets for others to train (and especially evaluate) on, as mislabeled samples in test sets have been shown to destabilize ML benchmarks [3]. We will add some discussion on this to the revised version of the paper.

Finally, per-sample mislabel identification methods such as LEMoN can also be deployed to flag incorrect human inputs in an online setting. One particular example might be flagging simple mistakes made by radiologists when writing notes from chest X-rays (as motivated by our MIMIC-CXR setting).

[1] Learning with Noisy Labels Revisited: A Study Using Real-World Human Annotations. ICLR 2022.

[2] Beyond Synthetic Noise: Deep Learning on Controlled Noisy Labels. ICML 2020.

[3] Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks. NeurIPS 2021.

Thank you for the insightful review and constructive feedback!

For instance, BLIP-based filtering methods may not be prohibitively expensive, as they can be fine-tuned on a domain-specific dataset (e.g., movie or biomedical domains) with relatively modest computational costs. Such fine-tuning could potentially outperform the proposed method—as suggested by results on the MS-COCO dataset. A direct comparison of computational overhead between LEMoN and these BLIP-based approaches would strengthen the paper’s claims. Furthermore, CapFilt appears to perform reasonably well even in a zero-shot setting (i.e., without fine-tuning on MS-COCO), wondering about this simple zero-shot CapFilt result. Clarifying the necessity of LEMoN in contrast to these simpler alternatives would help reinforce its practical relevance.

CapFilt Inference Runtime: We compare the inference time (per sample runtime, milliseconds) of LEMoN with CapFilt using the same setup as Table I.11. We find that the two methods have generally comparable inference runtimes.

Advantages of LEMoN over CapFilt: We clarify that for the results reported in the paper, we utilized the pretrained “Salesforce/blip-itm-base-coco” checkpoint. This model was trained on the clean training split of MSCOCO, which is why we refer to it as the "oracle". This training includes minimizing image-text matching loss, which is a binary classification objective designed to predict whether a caption matches a given image. LEMoN does not ever need access to the clean dataset (except optionally for hyperparameter tuning), only the noisy data. As such, we clarify that CapFilt is not applied in a “zero shot” setting, especially not for MSCOCO.

Further, we note that CapFilt appears to be more domain specific than LEMoN. As CapFilt has been trained on clean MSCOCO, it does well on MSCOCO and Flickr30k (both contain COCO‐style captions), but does worse than LEMoN on mmimdb.

At first glance, LEMoN's strong performance on mimiccxr seems to highlight its generalizability and ease of adaptation to out-of-domain data—a potential advantage over BLIP-based methods. However, since the proposed method also relies on a domain-specific encoder (BiomedCLIP), the comparison may not be entirely fair. I believe the advantage over the BLIP variant should be more clearly explained —particularly through comparisons with zero-shot CapFilt and fine-tuned CapFilt (e.g., on domain-specific datasets), along with an analysis of their relative computational costs. In addition, exploring whether the proposed scoring function could be integrated into or combined with existing methods like CapFilt might potentially enhance its practical utility.

We clarify that all other baselines on MIMIC-CXR except CapFilt do also utilize BiomedCLIP (where applicable), and LEMoN outperforms all such baselines. We highlight that we have also explored label error detection without an external domain-specific encoder for MIMIC-CXR in Table 4. Finally, as the reviewer points out, one can leverage representations from BLIP for label error detection with LEMoN, and LEMoN's score could also be combined with other mislabel scores (e.g. through ensembling). We highlight this as one area of future work.

Finally, many recent vision-language pipelines rely on synthetically generated high-quality captions (e.g., post-BLIP processing). In such scenarios, the role and added benefit of LEMoN seems less obvious.

We highlight that such vision-language pipelines assume that high quality synthetic caption generators already exist. LEMoN is designed to improve these synthetic caption generators by providing them with better real training data to start with (e.g. Section 6.2). Additionally, LEMoN may be used to detect errors in synthetic captions as well, thus adding the potential of improving the filtering component of such vision-language pipelines.

believe the paper includes relevant citations overall; however, the main topic is also closely related to the issue of false negatives in vision-language models. The authors may consider to include references such as [Chun et al., ECCV 2022] and [Byun et al., MAFA, CVPR 2024]

Thank you for these references – we will add them in the revised version of the paper!