The Pursuit of Human Labeling: A New Perspective on Unsupervised Learning

We present HUME, a simple model-agnostic framework for inferring human labeling of a given dataset without any external supervision. The key insight behind our approach is that classes defined by many human labelings are linearly separable regardless of the representation space used to represent a dataset. HUME utilizes this insight to guide the search over all possible labelings of a dataset to discover an underlying human labeling. We show that the proposed optimization objective is strikingly well-correlated with the ground truth labeling of the dataset. In effect, we only train linear classifiers on top of pretrained representations that remain fixed during training, making our framework compatible with any large pretrained and self-supervised model. Despite its simplicity, HUME outperforms a supervised linear classifier on top of self-supervised representations on the STL-10 dataset by a large margin and achieves comparable performance on the CIFAR-10 dataset. Compared to the existing unsupervised baselines, HUME achieves state-of-the-art performance on four benchmark image classification datasets including the large-scale ImageNet-1000 dataset. Altogether, our work provides a fundamentally new view to tackle unsupervised learning by searching for consistent labelings between different representation spaces.

Paper

References (84)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer BUAb7/10 · confidence 3/52023-07-03

Summary

The paper proposes a framework, called HUME, for providing human labels without any supervision. The method uses an assumption that classes derived from human labelling are linearly separable regardless of the representation space used to represent the dataset, ie invariant to sufficiently strong representation spaces. The authors show that this method generates labels that are correlated with human labeling (ie ground truth) of the dataset. HUME also shows good performance on STL-10 (ie better than a supervised linear classifier), and comparable performances on other datasets. Results are even more striking, for all the datasets used, when compared against other unsupervised methods. Finally, HUME provides reliable samples on CIFAR-10.

Strengths

From the related work explained in the paper, this is the first work in the literature to provide a method that does not rely on semantic knowledge of samples, eg clustering, but rather it uses a novel generalization perspective. The idea of using two pretrained models to generate human-like labels under an unsupervised manner is simple and interesting. The text is clear and overall well-written. The related work section is well expanded.

Weaknesses

The paper could improve the analysis on the correlation of their labels with the ones humans provide, ie it’s not clear whether it generalizes to other datasets/settings, and this is a very important validation experiment of the proposed method since it’s part of the hypothesis of the paper (ie that human labeled tasks are linearly separable in a sufficiently strong representation space). Another analysis that could be added to the appendix is on the reliability of samples for other datasets (eg expands Fig. 3a of CIFAR-10 for other datasets). For now, it is only shown to work for CIFAR-10. Minor typo: line 107 “we only train linear classifiers on top pretrained representations” should probably have an “of” between “top” and “pretrained” words.

Questions

* What do the authors mean by “unbiased estimate” (line 90)? * Which distance measure is used for fig. 1? Could this be clarified in the paper? * Could the authors provide in appendix the same fig. 1 but for other datasets (eg CIFAR-100-20 and STL-10)? Since being well-correlated with humans is essential for the proposed method, it would be interesting to see if this result generalizes for other datasets as well. * In the paper it is mentioned in both Definitions 1 and 2 the sentence “attains low test error” (lines 101 and 105), but how low is not commented throughout the text and this is left a bit uncommented… could the authors expand on that? Could the authors also provide some analysis/experiments on this front? * Could the authors provide a better explanation for the difference between inductive and transductive procedures? Is it the case that the inductive one a model is learned during training, and then just the model is used to make predictions on unseen data, and in the transductive case the data is kept for the inference as well (eg using some form of nearest neighbor analysis)? * In Table 1, could the authors provide an intuition behind why the results seem to improve more for the transductive than for the inductive procedure?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Yes, limitations have been well addressed.

Reviewer BH9W8/10 · confidence 3/52023-07-03

Summary

This paper proposes a novel unsupervised learning algorithm called HUME. HUME aims to minimize the disagreement between two linear models. The two linear models consume features from different backbones, and thus they are iteratively optimized to map difference feature spaces into the same label space. HUME has achieved very promising results on extensive experiments conducted in this paper.

Strengths

(1) This paper provides a fundamentally new view to tackle unsupervised learning by searching for consistent labelings between different representation spaces, i.e., human-labeled tasks are invariant to sufficiently strong representation spaces, it gives valuable insights to the following research works (2) HUME outperforms existing state-of-the-art unsupervised learning methods by a large margin and can achieve comparable or better performances than supervised learning on some tasks (3) HUME can produce reliable labels for semi-supervised learning to further boost the performance, which provides a good starting point for applications in many low-resource scenarios

Weaknesses

I don't find substantial weakness from this paper.

Questions

I am personally interested in applying semi-supervised learning to the reliable pseudo labels. Could you provide the following results? (1) Comparison of using reliable pseudo labels versus human labels (2) Results on more datasets This is not obligatory, and you could put the results in the appendix.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

The authors have raised 2 limitations and proposed reasonable solutions for mitigation. I don't find other limitations of this work.

Reviewer J4o25/10 · confidence 2/52023-07-05

Summary

This work proposes HUME, an unsupervised framework for inferring labels of an image dataset. Based on the assumptions of linear separability and model invariance, the authors propose an objective that utilizes two self-supervised models, one on the target dataset and another using large-scale pre-training. Specifically, a linear mapping is applied on the first model representation to generate pseudo labels, which is updated by MAML to minimize classification loss of another linear classifier on the second representation. HUME is empirically evaluated on three image datasets, outperforming deep clustering and sometimes also linear probing baselines.

Strengths

- The paper is well organized, with the entire model built around two key assumptions (linear separability and model invariance). Strong justifications are provided for each component. - Experiments are conducted on multiple dataset and models to validate the generalizability of HUME. - Ablations studies are included on aggregation over n runs, regularization weight, self-supervised representation etc.

Weaknesses

- My biggest concern is on the assumption of linear separability on pretrained representations. This is probably true for easy-to-classify datasets (e.g. CIFAR-10) but less likely for fine-grained tasks (e.g. CIFAR-100). There exists much more challenging vision datasets than the ones studied in the paper, where it may be difficult, if at all possible, to find a feature space that linearly separates all classes. I doubt HUME will scale favorably to these datasets. - Results in table 1 seems to confirm this suspicion, as the relative performance of HUME to linear probing drops quickly as the datasets becomes more difficult. - One may further argue that if the classes are already linearly separable, then only a few labeled examples per class are needed to achieve high accuracy, which should not cost substantial human effort to annotate. The most demanding tasks to annotate are often the ones for which a good representation does not exist. - The fairness of experimental comparisons is questionable: - HUME makes use of large pretrained models such as DINO, unlike other unsupervised methods that are learned on target dataset only. I am not sure to what extent the improvement should be attributed to the stronger feature representation of large models. The authors can consider adding comparison to deep clustering or linear probing using large models as backbone. - The authors compared to a variant of SPICE without joint training ($\rm SPICE_S$). The standard SPICE outperforms HUME by jointly updating the feature model. - If the class vocabulary is known, this also enables zero-shot classification by vision-language models (e.g. CLIP). I am curious how fully unsupervised methods like HUME compares to CLIP in assigning pseudo labels, or if the two can be combined into a stronger model (not just using CLIP ViT as a feature extractor, but facilitating its joint embedding space for image and text). - Despite the claim that HUME can benefit semi-supervised learning, no SSL models were trained and evaluated to verify this. The accuracy of reliable samples serves as an indirect proof, but cannot replace full semi-supervised experiments, e.g., by following the joint training step in SPICE.

Questions

Using 300 iterations (supp. L564) for the inner loop of MAML appears quite impractical. I would appreciate if the authors can comment on the computational cost of HUME, or correct any misunderstanding I had with the implementation.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations are discussed in the paper. There is no discussion of broader societal impacts as the paper focus on foundational learning algorithms; though the authors may consider discussing the possibility of biases of large pretrained models being transferred to the pseudo labels through HUME, etc.

Reviewer VWGE7/10 · confidence 3/52023-07-06

Summary

The work presents a new method for inferring human labeling without supervision. The main idea of the method is that human labels for a dataset should be linearly separable for all good representations, and relatively invariant to the representation space. The method uses bi-level optimization with an inner loop that optimizes for the pseudo-labels of a training set, and an outer loop which optimizes for the test loss of a classifier trained on the pseudo-labels. HUME achieves state-of-the-art performance on standard unsupervised learning benchmarks, CIFAR and STL-10.

Strengths

- The insight that human labels should be linearly separable in any sufficiently strong representation space is original and interesting. Showing that the objective function proposed in HUME is well correlated with human labeling is surprising, and would likely be useful to future work. - HUME outperforms state-of-the-art on standard unsupervised learning benchmarks by a fair margin and impressively outperforms the supervised linear probe on STL-10. - The paper is well organized and the writing is clear.

Weaknesses

HUME uses DINO and MOCO while other methods only use MOCO. It could be that DINO is a better representation given that HUME with DINO outperforms HUME with other representations. It’s also unclear whether the performance gain is coming from ensembling multiple models or the optimization procedure of HUME. It would be useful to see an ablation where HUME is trained only on a MOCO representation for a straight-forward comparison to other unsupervised methods and to understand how much each component of HUME contributes. As stated in the limitations, the number of classes and label distribution must be known a priori. Other works in this subfield make the same assumption so this weakness is not particular to HUME.

Questions

- Can the method scale to larger datasets considering that second-order gradients need to be calculated? Does approximating the Hessian affect performance? - MAML is known to be difficult to optimize with respect to hyper-parameters. Does this make HUME sensitive to hyper-parameters such as the number of inner-loop steps, learning rate, etc? - In equation 3, you constrain the weights of the linear layer, $W_1$, to be orthonormal, but I didn’t see a mention of the method you’re using to do so. What is the reasoning for this design choice?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations are well addressed by the authors.

Reviewer uZTE8/10 · confidence 4/52023-07-07

Summary

The authors propose a clever method that exploits recent findings about the effectiveness of linear probing across diverse representation spaces in order to discover semantically-meaningful clusters (i.e. ones that correspond to human labeling) by seeing which linearly-separable clusters are preserved across multiple representation spaces.

Strengths

- HUME is a very clever method that beautifully exploits recent findings about the effectiveness of linear probing across diverse representation spaces. - The paper is well-written and easy to follow.

Weaknesses

- It's unclear whether performance gains are driven by HUME or by the strength of the second representation space. Additional baseline experiments with linear probing and clustering on top of the large pre-trained model representations (DINOv2, CLIP, ViT) would help clarify this. Another nice baseline evaluation that could help would be to simply concatenate the embeddings from the two representation spaces and repeat the linear probing and clustering evaluations on that.

Questions

- When producing reliable samples for SSL, what is the class balance? The usefulness is impacted significantly by the diversity of the samples HUME can produce. I see in Fig 3a) that samples are produced from each class, but is performance roughly the same across all classes? - Are all three of the large-pretrained models that you use pre-trained in unsupervised ways (e.g. none were ever actually trained on the ImageNet classes)? - Unrelated to the review/scoring and purely out of curiosity, why did you name your method HUME? Is it an acronym for something?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

See weaknesses and questions sections.

Reviewer uZTE2023-08-11

Thank you for the excellent response! My concerns have been resolved, and I have updated my score accordingly. On a side note, the name is great and the acronym design, while definitely a bit of a stretch ("labEling"), is well-justified by the motivation and not even the most egregious acronym I've seen this reviewing season :-)

Authorsrebuttal2023-08-11

We thank the reviewer very much for the prompt response and for raising the score. We are very happy that the reviewer finds our response effective. We are also happy to hear that the reviewer likes acronym of our method and the story behind it.

Reviewer VWGE2023-08-14

Response to Author Rebuttal

I thank the authors for their response. My concerns have been addressed and all details have been clarified. The insight of using linear separability across representations to derive human-like labels is interesting and I maintain my score of accept.

Authorsrebuttal2023-08-14

We thank the reviewer for the response and for keeping confidence in our work. We are glad that our clarifications effectively addressed reviewer’s concerns. We are also happy to hear that the reviewer finds the key idea of our work interesting and useful for future work in this field.

Reviewer J4o22023-08-18

Thank you for their detailed response! The additional results are encouraging and addressed my primary concerns, and I have updated my rating to borderline accept. I feel that some of the rebuttal experiments (ImageNet, DINO w/ clustering, SSL) should be expanded and included in the paper to make the arguments more convincing. I am still a bit concerned that unsupervised learning is less critical when strong representation spaces exist ("if the classes are already linearly separable, then only a few labeled examples per class are needed to achieve high accuracy, which should not cost substantial human effort to annotate" in original review), though I see the potential of HUME as a bridge between unsupervised and semi-supervised learning in eliminating the need to label even a few examples. I would love to know the authors' view on this. Overall I am cautiously optimistic that this work presents sufficient contributions to be accepted to NeurIPS.

Authorsrebuttal2023-08-18

> Thank you for their detailed response! The additional results are encouraging and addressed my primary concerns, and I have updated my rating to borderline accept. I feel that some of the rebuttal experiments (ImageNet, DINO w/ clustering, SSL) should be expanded and included in the paper to make the arguments more convincing. We thank the reviewer for the response and for raising the score. We are glad that our clarifications helped to increase the reviewer's confidence in our work. We will include these additional experiments in our paper and we thank the reviewer for the suggestions. Below we address the remaining concern of the reviewer. We hope that our response will further help to increase the reviewer's confidence in our work. >I am still a bit concerned that unsupervised learning is less critical when strong representation spaces exist ("if the classes are already linearly separable, then only a few labeled examples per class are needed to achieve high accuracy, which should not cost substantial human effort to annotate" in original review), though I see the potential of HUME as a bridge between unsupervised and semi-supervised learning in eliminating the need to label even a few examples. I would love to know the authors' view on this. We thank the reviewer for an additional question. In response to the reviewer’s feedback, we have conducted additional experiments that compare HUME to few-shot learning baselines. Namely, starting from the strongest DINO representation space, we employ two different few-shot baselines: (i) following [1], we compute a prototype for each class on top of DINO representations using few-labeled examples and then assign examples to the nearest prototype (we refer to it as proto baseline), and (ii) we train a linear classifier on top of DINO representations using a few labeled examples (we refer to it as linear baseline). Remarkably, the results show that HUME consistently outperforms these two baselines in the 1-shot and 2-shot learning settings on all datasets. In the 3-shot learning setting HUME outperforms these baselines on the STL-10 dataset, matches performance on the CIFAR-10 dataset, and these baselines show better performance only on the CIFAR100-20 which in this setting requires *60 labeled examples*. In particular, in the 1-shot learning setting HUME significantly outperforms proto baseline by 39%, 24% and 44% on the STL-10, CIFAR-10 and CIFAR-100-20 datasets, respectively. Similarly, HUME outperforms linear baseline by 39%, 23% and 37% on the STL-10, CIFAR-10 and CIFAR-100-20 datasets, respectively. In the 2-shot learning setting, HUME outperforms proto baseline by 18%, 7% and 6% on the STL-10, CIFAR-10 and CIFAR-100-20 datasets, and linear baseline by 22%, 8% and 4%, respectively. It is important to emphasize that in this setting these baselines use *20 labeled examples* for STL-10 and CIFAR-10, and *40 labeled examples* for CIFAR-100-20 while HUME works *fully unsupervised*. Only, in the 3-shot learning setting we start seeing the benefits of few-shot baselines but only on the CIFAR-100-20 dataset. In particular, on the STL-10 dataset HUME still outperforms proto and linear baselines by 9% and 11% respectively, and matches performance of these baselines on the CIFAR-10 dataset, and proto and linear baselines outperform HUME by 11% and 13% on CIFAR-100-20 dataset by using 60 labeled examples. On the STL-10 dataset, we observe that HUME keeps outperforming these baselines even in a challenging 5-shot learning setting where these baselines use 50 labeled examples. We would also like to emphasize that obtaining few-labeled examples per class is in reality hard because it requires a strong prior knowledge of knowing the complete space of all possible classes in advance and then labeling a few examples per class. In many real-world tasks, this is impractical and very difficult. We thank the reviewer for this question and we will also include few-shot baselines in our paper. We believe these results further strengthen our work. Overall, our results show that although being completely unsupervised, HUME (i) outperforms *fully supervised* linear model on top of MOCO representations on the STL-10 dataset and matches the performance on the CIFAR-10 dataset (main paper results); (ii) outperforms challenging *few-shot learning baselines* ran on top of DINO representations in the 1-shot and 2-shot learning scenarios, and (iii) *semi-supervised method FreeMatch* using reliable samples from HUME matches the performance of the FreeMatch with ground truth labels. [1] Snell et al. Prototypical Networks for Few-shot Learning. NeurIPS 2017.

Reviewer BUAb2023-08-19

Thanks authors for the detailed rebuttal. My concerns have been addressed, and I have updated my score accordingly.

Authorsrebuttal2023-08-19

We thank the reviewer very much for the response and for raising the score. We are happy that the reviewer finds our response effective.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC