LMC: Large Model Collaboration with Cross-assessment for Training-Free Open-Set Object Recognition

Summary

In this paper, the authors tackle the problem of open-set object recognition. The key challenge in open-set recognition problem they tackle is the model's reliance on spurious-discriminative features. To tackle this challenge, they propose an open-set object recognition system consisting of multiple pre-trained foundation models. They employ ChatGPT to generate virtual open-set categories that share spurious-discriminative features with closed-set categories. Then they generate images of the both closed-set and open-set categories by using DALL-E. During testing, they employ DINO and CLIP to get more accurate predictions by matching a test image to the generated images and, by matching a test image to category embeddings. They validate the proposed method on public benchmarks.

Strengths

This paper has the following strengths. S1: The problem tackled, open-set object recognition is interesting and challenging. S2: The proposed approach is reasonable. Using generative models to leverage knowledge-base and image generation capability could improve open-set recognition performance by mitigating spurious correlations between the category and features. S3: The proposed method shows favorable performance on public benchmarks, without any fine-tuning. S4: The ablation experiments show the effectiveness of the proposed design choices. Chain-of-thought reasoning, self-checking, and CLIP-based feedback all contribute to the final performance. Furthermore, the performance improvement is not merely from the category overlap between the close-set and the virtual open-set as shown in Table 6. The improvement might come from mitigated spurious correlation.

Weaknesses

This paper has the following weaknesses. W1: Although the proposed method is effective in open-set recognition, the technical contribution is somewhat limited. This is okay if the performance improvement compared to the baseline is very significant. However, the performance improvement compared to the baseline is not surprising given that the proposed method consists of four foundation models: GPT, DALL-E, CLIP, and DINO. For example, the proposed method shows 3.2 points AUROC improvement and 5.5 points OSCR improvement over the softmax baseline in TinyImageNet. W2: The authors claim the improved performance comes from mitigating the spurious correlation. However, there is no quantitative evaluation on the amount of spurious correlation mitigated. It would be better to show some quantitative evidence of how much the proposed method mitigates the spurious correlation.

Questions

Please address the weaknesses part. I have a few additional questions below. Q1: What is the performance if we learn a few parameters on top of the foundation models used? For example, what is the performance if we add some adapter layers on top of CLIP and/or DINO to fine-tune the model on the target datasets? Can we get an even stronger performance? Q2: Does the proposed method work well on more fine-grained recognition tasks? E.g., CUB or Oxford Flowers datasets. It seems that the proposed method heavily relies on ChatGPT’s common sense reasoning ability. What happens if ChatGPT cannot generate reasonable descriptions of certain fine-grained classes? Since the proposed method does not learn anything, it might fail if ChatGPT fails. I would like to listen to the authors' opinions and/or empirical validation on this issue.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Contribution

2 fair

Limitations

I could not find limitation and broader impact part.

Summary

This work demonstrates a very interesting and sophisticated usage of various large models, including ChatGPT, CLIP, DINO and DALL-E., to tackle the open-set recognition problem. The main challenge in open-set recognition is that the chosen classifier may get confused by certain spurious-discriminative features that are shared between the closed- and open-set classes. The main idea of this work is to leverage the text-to-image alignment capability of CLIP and image-to-image alignment capability of DINO with the aid of virtual open-set classes. The list of virtual open-set classes is elicited from LLM, i.e. ChatGPT, through carefully designed prompts that introduces intermediate reasoning and self-checking. Additionally, the descriptive text can then serve as prompts for DALL-E to synthesize the representative images of the corresponding open-set classes. During inference time, both the pre-trained CLIP and DINO models can then utilize the text class names and synthetic representative images to perform zero-shot (training-free) classification, and meanwhile reducing the influence of spurious-discriminative features. Evaluation on several benchmarks demonstrates new state-of-the-art performance.

Strengths

1. The paper is well-written and very easy to follow. The additional information provided in the supplemental material is also very helpful in resolving most of the doubts I had when reading the main manuscript. In the beginning, I particularly felt that how the large models collaborate and self-improve to accomplish their jobs was a little magical and I wanted some more transparency after reading the main manuscript. I appreciate that the authors had put sufficient information in the supplemental material to make the whole work more technically sound and convincing. 2. This work tackles the open-set recognition from the system-level and demonstrates a very clever usage of several off-the-shelf large models of different modalities. Although it didn't introduce new techniques to improve any of the models involved, it successfully combined them to achieve superior performance. From the practical point of view, this method possesses many advantages, such as being training free and fast inference, that are important in real-world applications.

Weaknesses

1. Although large model collaboration is an interesting way of solving open-set recognition problem, it treats the large models as black boxes as it is. and it is perhaps not easy to come up with an even more sophisticated method to build a LMC system like this work, In some sense, I feel that it does not pave a new way for follow-up research, and yet I do not believe the open-set recognition problem is completely solved by the LMC approach. 2. Although I don't think the evaluation of this work is insufficient, I think it is still worth validating the effectiveness of LMC framework on larger benchmarks such as the Semantic Shift Benchmark proposed in [48]. I doubt if the construction of virtual open-set classes by LLM will fail at a certain point as the number of fine-grained visually similar close-set classes grows. Another drawback of using LLM is that it is very difficult for human to "fact check" the answers and detect any content that is made up by LLMs.

Questions

From the exemplar dialogs with ChatGPT, the conversation still seems a bit open-ended and unstructured. I am curious about how the authors completed the preparation of the virtual open-set list, synthesis of diverse representative images and pre-compute the CLIP and DINO features in just about 38.9 minutes (Appendix, Table 9). I assume that this is done programmatically. Are there any tricks involved here to simplify the post-processing of ChatGPT answers? As far as I know, the most important tunable parameter when calling OpenAI API is the temperature, which controls the diversity of the results that the LLM generates. How did you set this parameter? Does it affect the quality of ChatGPT's self-checking process?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Contribution

3 good

Limitations

The performance of LMC framework is apparently affected by the large models and there is no discussion about it. I am particularly curious about the potential failure modes of ChatGPT.

Summary

This paper proposed to incorporate several large models to solve the spurious-discriminative problem in open-set recognition. It first prompts ChatGPT to describe the known classes and generate a list of new classes that share the spurious-discriminative features with the known classes. Then it prompts DALL-E to generate the images of known and unknown classes. It also uses CLIP to detect the less accurate generated images and asks ChatGPT to refine the description and uses DALL-E to generate again. During inference, it uses both the CLIP model and DINO model to calculate the uncertainty scores and fuse their results. The experiments and ablation study are comprehensive. Overall, I think this paper provides a new alternative solution for open-set recognition, and the data generation process could be very useful for the industry.

Strengths

1. The overall idea of using several large models for training-free open-set recognition is new and interesting. The pipeline of using ChatGPT to generate the new classes that share the spurious-discriminative features with the known classes is reasonable. The refinement procedure using CLIP and ChatGPT is very rigorous and effective. 2. The inference pipeline that fuses the results of CLIP and DINO is also reasonable. It is somewhat similar to the method in the few-shot open-set recognition [1, 2], which also fuses the uncertainty score from two aspects. The authors may consider citing the related papers. 3. The experiments and ablation studies are very comprehensive. Reference: [1] Ssd: A unified framework for self-supervised outlier detection. In ICLR, 2022. [2] THE DEVIL IS IN THE WRONGLY-CLASSIFIED SAMPLES: TOWARDS UNIFIED OPEN-SET RECOGNITION. In ICLR 2023.

Weaknesses

The main pipeline is a kind of data generation process, which could not be done online. So although it is a good try to use several large models and coordinate them to generate wanted data, it is more like an engineering project rather than a research topic.

Questions

1. How much time does the data generation cost for different datasets? 2. What if we use other pre-trained image models instead of DINO? Like ImageNet pre-trained or MoCo?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Contribution

3 good

Limitations

None.

Summary

This paper proposes an open-set recognition framework, Large Model Collaboration (LMC), which collaborates several large models (ChatGPT, DALL-E, CLIP, and DINO) to make use of the rich implicit knowledge to reduce the reliance on spurious-discriminative features. The proposed framework consists of the following two stages: (i) The first stage is to simulate virtual open-set classes, including simulating names and generating images. In order to improve the effectiveness of the simulated names for open-set classes, ChatGPT is asked with three designed questions and the corresponding intermediate rationales for obtaining names of new classes that share spurious-discriminative features with each closed-set class, and is guided to perform iterative self-check (where the new simulated class names are used to expand the name list, and the expanded name list is used to obtain new simulated class names) for covering as many spurious-discriminative features as possible. Then, ChatGPT is asked to generate diverse descriptions for each class in the expanded name list, and DALL-E is used to generate images based on these descriptions. In order to improve the accuracy of the generated images, a cyclic self-assessing module is proposed, where ChatGPT is provided with feedbacks from CLIP about the less accurate images to refine the descriptions generated by ChatGPT and DALL-E is used to re-generate images based on the refined descriptions. At this stage, an expanded list of both closed-set and simulated open-set classes and diverse generated images for each class in the list are obtained. (ii) The second stage is for inference based on the expanded list and the generated images. At this stage, two alignments (image-text alignment by CLIP and image-image alignment by DINO) are performed for obtaining the corresponding scores to each testing image, and the weighted result of the two scores is used for open-set recognition.

Strengths

(1) The idea, which makes use of several large models for open-set recognition, is somewhat novel and interesting. (2) The writing is clear and easy to follow. Some examples are well visualized. (3) Experimental results on four small-scale datasets (CIFAR10, CIFAR+10, CIFAR+50, and TinyImageNet) demonstrate that the proposed methods performs better than the comparative methods.

Weaknesses

(1) Lacking comparisons under the cross-dataset setup: The proposed framework is only evaluated under the standard setup where the open-set-class images and closed-set-class images are from a same dataset. Such experiments are insufficient, and in fact, many open-set recognition methods [15,21,36,45,55,63] have also been evaluated under the cross-dataset setup. (2) Lacking comparisons on larger-scale datasets: in this paper, only four small-scale datasets (i.e., CIFAR10, CIFAR+10, CIFAR+50, and TinyImageNet) are used for evaluation. It is necessary to make comparisons on some larger-scale datasets, e.g., CUB and FGVC-Aircraft. (3) Comparison with other image generative methods: One main contribution of the proposed framework is generating virtual open-set-class images. Many other works (e.g., [8,21,30]) also generate virtual images, and it is better to replace the image generation method used in the proposed framework with some of them, and make a further comparison. (4) Lacking details of efficiency: One advantage of the training-free methods is its high efficiency. In the proposed framework, although there is no training stage, the pre-processing stage and the inference stage which is based on two alignments may also bring additional computational costs. Hence, it is suggested to provide details about the efficiency of the two stages. (5) Lacking discussion about the ablation results: In Sec 4.4, several ablation studies are conducted. However, the authors only stated that the final results were improved, but didn’t give any discussion on the effect of each module.

Questions

Please see the weaknesses above.

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

The authors did not adequately address the limitations. Two additional suggestions are listed here: (1) It’s better to add the full name of LMC, i.e., Large Model Collaboration, in the abstract in line 6. (2) In Line 219, it is suggested to add a brief description of 5, although Fig. 5 is also described in Line 415.

Response to the rebuttal

I carefully read the rebuttal and the reviews of other reviewers, and I believe that my concerns are solved. This work is valuable for the research community and industry, so I raise my score to 6.

Response to authors' rebuttal

I appreciate the great efforts made by the authors to prepare the rebuttal. I was impressed by the superior performance shown in Semantic Shift Benchmark and other experiments in the responses to other reviews. The rebuttal addressed most of my concerns and questions. I still hold my opinion that the LMC framework leaves less room for future exploration because large models remain a black box and attempts in (1), (3) of A1 could be easily regarded as *tricks* or lack of technical contribution. Nevertheless, I am in favor of accepting this paper and will keep my rating as Weak Accept.

Acknowledging the rebuttal and other reviews

I have read the rebuttal and other reviews. The rebuttal from the authors resolved most of my concerns. I appreciate the clarification on the softmax baseline and the additional analysis on the spurious correlation. The proposed method seems to have merit. I am leaning toward accepting this paper. I am increasing my rating to 6.