Robustness diagnosis method for fine-grained perception capability of visual language large model

Through the use of the Hasani dividend metric, the language and visual synergy of the visual language model are systematically evaluated, which solves the problem of insufficient optimization of the combination reasoning ability of the visual language model, and achieves in-depth analysis and performance improvement of the internal representation of the visual language model.

CN119027962AInactive Publication Date: 2024-11-26ZHEJIANG LAB
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202411516583.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-29
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing technology lacks methods that can deeply analyze the internal representation of visual language models from a theoretical basis, resulting in insufficient optimization of the combination reasoning ability of visual language models and ineffective improvement of model performance.

Method used

Using the measurement criteria of Hasani dividend, systematic evaluation is carried out from the perspective of language, vision, language and vision coordination. Through step one, the semantic interaction relationship is diagnosed, step two, the visual object's differential perception ability is evaluated, and step three combines the evaluation of visual encoder and text The encoder's object attribute combination relationship is used to quantitatively evaluate the fine-grained perception ability of visual language models.

Benefits of technology

Provide an interpretable diagnostic method for combining reasoning ability of visual language models, revealing the fragility of the model in combinatorial understanding, propose new metrics from the mathematical level, systematically analyze the internal representation of visual language models, which is better than existing methods, It is universal and explainable.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027962A_ABST
    Figure CN119027962A_ABST
Patent Text Reader

Abstract

The invention discloses a robustness diagnosis method for the fine-grained perception capability of a visual language large model, and the method carries out the systematic evaluation and analysis of the internal representation of the visual language model from the three perspectives of language, vision, and language and vision collaboration based on the mathematical basic theories, such as a Shapril value. And the internal mechanism of the robust reasoning capability defect is disclosed. The method mainly comprises the following steps of: explaining and evaluating whether a text encoder can understand a subtle change relationship of a text mode; explaining and evaluating whether the visual encoder can perceive the fine-grained difference between the image targets; the synergy of the visual encoder and the text encoder is interpreted and evaluated. The universal visual language large model robust reasoning ability defect diagnosis method is provided for the first time, new measurement is provided from the mathematical level to evaluate the vulnerability of the large model in different aspects of combined understanding, systematic analysis is conducted on internal characterization of the visual language model, and the method has the advantages of universality, interpretability and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a robustness diagnosis method for fine-grained perception capability of a large visual language model. Background Art

[0002] The Shapley value (L. Shapley. 1953. A Value for n-Person Games. Contributions to the Theory of Games, 2:307-317.) was originally proposed in game theory and is used to fairly distribute the total rewards obtained in a working alliance. It is the only unbiased estimation metric. The Shapley value satisfies the linearity, imaginary, symmetric and efficiency properties (E.Roth, LS Shapley. 1991. The Shapley Value: Essays in Honor of Lloyd S.Shapley. Economica, 101-123.). Based on the above characteristics, a large number of studies currently use the Shapley value to measure the interaction between input variables, and the interaction based on the Shapley value has also been widely used in various fields of DNNs. (SM Lundberg, S.-I. Lee. 2017. A Unified Approach to Interpreting Model Pre-dictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems, ACM, 4768-4777.) defines the Shapley additive interpretation interaction value and gives the prediction explanation of tree ensemble methods such as XGBoost and LightGBM (SM Lundberg, GG Erion, S.-I. Lee, et al. 2018. Consistent Individualized Feature Attribution for Tree Ensembles. arXiv preprint, arXiv:1802.03888.).(M. Tsang, D. Cheng, H.Liu, et al. 2020. Feature Interaction Interpretability: A Case for Explaining Ad-Recommendation Systems via Neural Interaction Detection. In Proceedings of the 8th International Conference on Learning Representations, ICLR.) proposed a statistical interaction detection framework for explaining neural networks (M. Tsang, D. Cheng, Y. Liu. 2018. Detecting Statistical Interactions from Neural Network Weights. In Proceedings of the6th International Conference on Learning Representations, ICLR.) and provided an explanation for the recommendation model. (M. Ancona, C. Oztireli, M. Gross. 2019. Explaining Deep Neural Networks with a Polynomial Time Algorithm for Shapley Values ​​Approximation. In Proceedings of the 36th International Conference on Machine Learning, ACM, 272-281.) A polynomial time complexity calculation method is proposed to solve the problem of high time complexity of calculating Shapley values. (K. Dhamdhere, A. Agarwal, M. Sundararajan. 2019. The Shapley Taylor Inter action Index. In Proceedings of the 36th International Conference on Machine Learning, ACM, 9259-9268.) The Shapley-Taylor index is proposed, which decomposes the prediction of DNNs into the interaction effects of feature subsets.Zhang et al. (H. Zhang, X. Cheng, Y. Chen, et al. 2020. Game-Theoretic Interactions of Different Orders. arXiv preprint, arXiv:2010.14978.) defined multivariate interactions and multi-order interactions to explain DNNs (H. Zhang, Y. Xie, L.Zheng, et al. 2021. Interpreting Multivariate Shapley Interactions in DNNs. In Proceedings of the 35th AAAI Conference on Artificial Intelligence, AAAI, 10877-10886.). (H. Deng, Q. Ren, H. Zhang, et al. 2022. Discovering and Explaining the Representation Bottleneck of DNNs. In Proceedings of the 10th International Conference on Learning Representations, ICLR.) used multi-order interactions to explore the representation bottleneck of DNNs. However, to the best of our knowledge, no one has yet used game theory to diagnose internal representation deficiencies in vision-language models and use this as a basis for optimizing the model's robust reasoning capabilities.

[0003] Compositionality plays an integral role in vision and language. Previous work exploring combined reasoning of vision and language covers a variety of tasks such as visual question answering (C. Saharia, W. Chan, S. Saxena, et al. 2022. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In Proceedings of the 36th International Conference on Neural Information Processing Systems, ACM, 36479-36494.), generation (Ramesh, P. Dhariwal, A. Nichol, et al. 2022. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint, arXiv:2204.06125.), retrieval, and planning (Conwell, T. Ullman. 2022. Testing Relational Understanding in Text-Guided Image Generation. arXiv preprint, arXiv:2208.00005.). Combinatorial reasoning capability refers to the model’s ability to understand complex scenes and texts, including distinguishing sentences with the same words but in different orders, and scenes with the same objects but in different colors.Previous studies have designed a series of new benchmarks to comprehensively evaluate the combinatorial reasoning ability of visual language. For example, Winoground (T. Thrush, R. Jiang, M. Bartolo, et al. 2022. Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality. 2022 IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 5228-5238.) tests the performance of the model on rich text sequential structures, CREPE (Z. Ma, J. Hong, MO Gul, et al. 2023. CREPE: Can Vision-Language Foundation Models Reason Compositionally. 2023 IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 10910-10921.) tests the performance of the model on the composition of objects, their relationships and attributes, and the ARO benchmark (M. Yuksekgonul, F. Bianchi, P. Kalluri, et al. 2022. When and why vision-language models behave like bags-of-words, and what to do about it. InProceedings of the 10th International Conference on Learning Representations,ICLR.) is a benchmark that systematically evaluates the ability of models to understand different types of relations, attributes, and order. VL-Checklist (T. Zhao, T. Zhang, M. Zhu, et al. 2023. VL-CheckList: EvaluatingPre-trained Vision-Language Models with Objects, Attributes and Relations.arXiv preprint, arXiv:2207.00221.) is a framework for evaluating the ability of models to recognize objects, attributes, and relations.Cola (Ray, F. Radenovic, A. Dubey, et al. 2023. Cola: A Benchmark for Compositional Text-to-image Retrieval. In Proceedings of the 37th International Conference on Neural Information Processing Systems, ACM, 46433–46445.) analyzes the model's combination capabilities in detail and proposes a text-to-image retrieval benchmark.

[0004] The above research mainly focuses on two aspects: 1. Evaluate the combinatorial reasoning ability of the visual language model in a black-box detection manner (DH Park, S. Azadi, X.Liu, et al. 2021. Benchmark for Compositional Text-to-Image Synthesis. arXivpreprint, arXiv:2312.02338.), usually measuring the accuracy of whether the model correctly retrieves the matching text for a given image between two titles (C.-Y. Hsieh, J. Zhang,Z. Ma, et al. 2023. SugarCrepe: Fixing Hackable Benchmarks for Vision-Language Compositionality. In Proceedings of the 37th International Conference on Neural Information Processing Systems, ACM, 31096-31116.). 2. Improve the compositionality of the model in an empirical way by introducing supervision of the scene graph. However, these two methods lack mathematical connections with classic interpretable interaction indicators, and therefore cannot conduct in-depth analysis of the internal component representations of the visual-language model. This leads to a lack of theoretically based combinatorial reasoning optimization methods and an inability to improve model performance from fundamental principles. Summary of the invention

[0005] The purpose of the present invention is to provide a robust diagnosis method for the fine-grained perception ability of a large visual language model in view of the deficiencies in the prior art.

[0006] The objective of the present invention is achieved through the following technical solution: A robustness diagnosis method for fine-grained perception capability of a large visual language model, comprising the following steps:

[0007] Step 1: Diagnose the semantic interaction relationship from the perspective of language entities and evaluate the text encoder's ability to understand the changing relationship between text patterns;

[0008] Step 2: Evaluate the fine-grained difference perception ability under the visual object perspective to evaluate the visual encoder's ability to understand the changing relationship between different visual targets;

[0009] Step 3: Combine the evaluation results of the collaboration between the language entity perspective and the visual object perspective to evaluate the understanding capabilities of the visual encoder and the text encoder for the collaborative object attribute combination relationship.

[0010] Furthermore, the step 1 is specifically as follows: under the premise of keeping the text entities unchanged but changing the combination method of the text entities, the Hassani dividend metric is used to quantitatively evaluate the sensitivity of the text encoder to changes in the text pattern, and to evaluate the text encoder's ability to understand the combination relationship between text entities.

[0011] Furthermore, the step 1 includes the following sub-steps:

[0012] 1.1. Given a set of image-text pairs , by exchanging Get the perturbation text from the entity words in ; Align image text and Input the visual language model separately to obtain two matching scores: and ;in, represents the original image, Represents the original text, represents the perturbed text obtained by exchanging entity words, Represents all visual concepts on the original image, represents all the words in the original text, represents all the words in the perturbation text;

[0013] 1.2. Use the matching score obtained in step 1.1 and , in a given image and words , Based on different subsets of Quantitatively measure the sensitivity of text encoders to text patterns; represents a subset of words in the original text, represents a subset of words in the perturbed text;

[0014] 1.3. Using the matching score obtained in step 1.1 and , calculate the reward difference and interaction effect differences , and then calculate the word and Pearson correlation coefficient for different subsets of , and , thereby further testing the combinatorial reasoning ability of the text encoder; among them, Indicates when and The difference in the interaction effect only includes the object words, Indicates when and The difference in the interaction effect only includes the relational terms, Indicates when and The differences in the interactive effects of both relation words and object words are included.

[0015] Furthermore, in step 1.2, the sensitivity assessment includes: Sensitivity assessment, attribute words Sensitivity assessment, target words Sensitivity assessment, relational words and object words Sensitivity assessment of interactions and attribute and object terms Sensitivity assessment of interactions.

[0016] Furthermore, step 2 is specifically as follows: given two images containing the same object but with significantly different relationships between the objects; quantifying the change in the causal effect of the visual encoder output, analyzing the sensitivity of the visual encoder to changes in the relationship between different object components; and evaluating the visual encoder's ability to understand the combined relationship between visual targets.

[0017] Furthermore, the step 2 includes the following sub-steps:

[0018] 2.1. Given a set of image-text pairs , by changing The entity interaction relationship in obtains the perturbed image ; Align image text and Input the visual language model separately to obtain two matching scores: and ;in, represents the original image, Indicates changing the entity interaction relationship of the original image to obtain the perturbed image. Represents the original text, Represents all visual concepts on the original image, represents all the words in the original text, Represent all visual concepts on the perturbed image;

[0019] 2.2. Use the matching score obtained in step 2.1 and , in a given text and visual concepts and Based on different subsets of To quantitatively analyze the sensitivity of the visual encoder to two types of fine-grained differences between image objects: Sensitivity evaluation and paired objects The sensitivity assessment of the relationship between them is used to diagnose the visual encoder's ability to understand visual components; among them, A subset of visual concepts representing the original image, A subset of visual concepts representing the perturbed image, express and Each has its own object area, express and Two object areas at the same time;

[0020] 2.3. Use the matching score obtained in step 2.1 and , calculate the reward difference and interaction effect differences , and then calculate the visual concept and Pearson correlation coefficient for different subsets of and , to further test the combinatorial reasoning ability of the visual encoder; among them, express and The difference in the interaction effect of the object areas they each have, express and The difference in the interactive effect of two object regions simultaneously.

[0021] Furthermore, the step 3 is specifically as follows: based on the interpretation and evaluation results of the steps 1 and 2, further measuring the correlation between the specific interaction relationship of the image entities and the specific combination relationship of the text entities, so as to quantitatively evaluate the synergistic effect of the visual encoder and the text encoder.

[0022] Furthermore, the step 3 includes the following sub-steps:

[0023] 3.1. Evaluating the function based on sensitivity , measuring the influence of specific components of the image on the interaction between text object words and relation words , to explain when only given When the image area of ​​the neutron object is and How will the interaction between object words and relation words in the text affect the output of the visual language model? represents the original image, Represents the original text, represents the perturbed text obtained by exchanging entity words, Indicates a fixed original image Sub-image area, change the original text object words and relation words;

[0024] 3.2. Evaluating the function based on sensitivity , measuring the influence of specific components of text on the interaction between image objects , to explain when only given When the object word in and How changes in the relationship between objects in the visual language model affect the output; represents the original image, Indicates changing the entity interaction relationship of the original image to obtain the perturbed image. Represents the original text, Indicates fixed original text The object word changes the original image object.

[0025] The present invention also provides an electronic device, comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the robustness diagnosis method of the fine-grained perception capability of a large visual language model.

[0026] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the program is executed by a processor, the method for robustly diagnosing the fine-grained perception capability of a large visual language model is implemented.

[0027] The beneficial effects of the present invention are: based on the basic mathematical theories such as Hassani Dividend, and guided by the overall architecture of the visual language model, the sensitivity of the visual-language model in the fine-grained understanding task is defined from a theoretical level, and the internal representation of the visual language model is systematically evaluated and analyzed from the three perspectives of language, vision, and language and vision synergy, proving and explaining the unique properties of the visual language model for fine-grained understanding tasks, and revealing the internal mechanism of the visual language model's combinatorial reasoning ability defects in an interpretable way. The present invention proposes for the first time an interpretable diagnostic method for the combinatorial reasoning defects of the visual language model, proposes a new metric from a mathematical level to evaluate the model's vulnerability in different aspects of combinatorial understanding, and systematically analyzes the internal representation of the visual language model, which is superior to existing methods and has the advantages of universality and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 It is a flow chart of the present invention;

[0030] Figure 2 It is a comparison diagram of the original text and the disturbed text of the present invention;

[0031] Figure 3 It is the relative word of the present invention Sensitivity assessment, target words Sensitivity assessment, relational words and object words Figure 1 shows a specific example of interaction sensitivity assessment;

[0032] Figure 4 1 is a comparison diagram of an original image and a perturbed image of the present invention, (a) is a schematic diagram of an original image, and (b) is a schematic diagram of a perturbed image obtained by changing the entity interaction relationship of the original image;

[0033] Figure 5 Each object of the present invention Sensitivity Assessment and Paired Objects A specific example diagram of the sensitivity assessment of the relationship between them. DETAILED DESCRIPTION

[0034] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0035] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0036] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0037] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0038] The core technology of this invention is to use the Hassani dividend metric to systematically evaluate and analyze the internal representation of the visual language model from three perspectives: language, vision, and language and vision synergy, so as to diagnose the defects of the combined reasoning ability of the visual language model in an interpretable way.

[0039] like Figure 1 As shown, the embodiment of the present invention proposes a robustness diagnosis method for fine-grained perception capability of a large visual language model, comprising the following steps:

[0040] Step 1: Using the principle of Hassani Dividend, we design an evaluation metric with a mathematical basis to quantitatively evaluate the sensitivity of the text encoder to different text combination patterns. Hassani Dividend is a typical metric in game theory that measures the interaction between a group of players. Specifically, given a group of players Play a game , you can get certain rewards. Here, Represents a function that converts the player Any subset of is mapped to a real number, representing the numerical reward obtained. In such a game, each player usually does not contribute to the reward alone, but interacts with each other, forming different alliances, which have an accidental impact on the final result. Based on the evaluation results, the degree of understanding of the complex semantic interaction relationship between entities by the text encoder is studied; specifically, the following sub-steps are included:

[0041] 1.1、Generation of matching scores: Figure 2 As shown, given a set of image-text pairs , by exchanging Get the perturbation text from the entity words in . and Input the visual language model separately to obtain two matching scores: and ,in, represents the original image, Represents the original text, represents the perturbed text obtained by exchanging entity words, Represents all visual concepts on the original image, represents all the words in the original text, represents all the words in the perturbation text, Representation of image The normalized vector, This article represents Normalized vector. and The change between is a measure of the sensitivity of the text encoder to changes in mixed text patterns.

[0042] 1.2. Sensitivity assessment for text changes: Using the matching score obtained in step 1.1 and , using the sensitivity evaluation function based on the Hassani dividend Quantitatively measure the sensitivity of the text encoder to changes in text mode. The sensitivity evaluation function used is as follows:

[0043]

[0044] in, represents a subset of words in the original text, represents a subset of words in the perturbed text, Represents all visual concepts on the original image, represents the normalization coefficient, Represents the contingency effect of a specific set of visual concepts and a specific set of words.

[0045] in, is the normalization coefficient, represents the mean of the chance effects produced by a particular set of words, Indicates that the selected word set comes from the original text and the perturbation text, represents a subset of the selected word set, represents the incidental effect of all visual concepts of the original image and a subset of the selected word set. Figure 3 As shown, in the given image and words and Under the premise of different subsets of , the sensitivity evaluation function is used to quantitatively measure the sensitivity of the text encoder to five types of changes in the text mode:

[0046] Relational words Sensitivity evaluation. This method measures the causal effect of relational words on the output of the VLM text encoder, where and Contains only relative terms.

[0047] Attribute word This method measures the causal effect of attribute words on the text encoder output of VLM, where and Contains only attribute words.

[0048] Object word The method measures the causal influence of the target word on the text encoder output of the VLM, where and Contains only the object word.

[0049] Relational terms and object terms Interaction sensitivity measure. This method measures the change in the causal effect of the interaction between the relation word and the object word, where and Contains both relational and object terms.

[0050] Attribute words and object words Interaction sensitivity measure. This method measures the change in the causal effect of the interaction between attribute words and object words, where and Contains both attribute words and object words.

[0051] (1.3) Test the combinatorial reasoning ability of the text encoder: Use the matching score obtained in step (1.1) and , calculate the reward difference and interaction effect differences , and then calculate the word and Pearson correlation coefficient for different subsets of , and ,in Indicates when and The change in causal effect when only the object word is included, express and The change in causal effect when only the relational terms are included, Indicates when and When both object words and relation words are included, the causal effect changes, which further tests the combinatorial reasoning ability of the text encoder.

[0052] Step 2: Explain and evaluate whether the visual encoder can perceive fine-grained differences between image objects: Given two images containing the same objects but with significantly different relationships between the objects, quantify the changes in the causal effect of the visual encoder output, analyze the sensitivity of the visual encoder to changes in the relationship between different object components, and study the degree of understanding of the visual encoder for fine-grained differences between image objects based on the evaluation results; specifically, it includes the following sub-steps:

[0053] 2.1、Generation of matching scores: Figure 4 As shown, given a set of image-text pairs , by changing The entity interaction relationship in obtains the perturbed image . and Input the visual language model separately to obtain two matching scores: and .in, represents the original image, Indicates changing the entity interaction relationship of the original image to obtain the perturbed image. Represents the original text, Represents all visual concepts on the original image, represents all the words in the original text, represents all visual concepts on the perturbed image, Representation of image The normalized vector, This article represents Normalized vector. and The change between is a measure of the sensitivity of the visual encoder to fine-grained differences between image objects.

[0054] 2.2. Sensitivity evaluation for image changes: Using the matching score obtained in step 2.1 and , using the sensitivity evaluation function based on the Hassani dividend To quantitatively analyze the sensitivity of the visual encoder to fine-grained differences between image targets, the sensitivity evaluation function used is as follows:

[0055]

[0056] in, A subset of visual concepts representing the original image, A subset of visual concepts representing the perturbed images.

[0057] in For normalization, indicates that the selected visual concept set comes from the original image and the perturbed image, and the indicator Measured when given text hour, and The detailed visual causal effects vary between Figure 5 As shown, in the given text and visual concepts and Under the premise of different subsets of , the sensitivity evaluation function is used to quantitatively measure the sensitivity of the visual encoder to two types of fine-grained differences between image targets:

[0058] Each object The method measures the causal effect of each individual object region on the output of the visual encoder of the vision-language model, where and The region containing objects of the same identity, i.e., objects .

[0059] Paired objects The method measures the sensitivity of the relationship between The causal effect of the relationship between and Contains two object areas.

[0060] 2.3. Test the combinatorial reasoning ability of the visual encoder: Use the matching score obtained in step 2.1 and , calculate the reward difference and interaction effect differences , and then calculate the visual concept and Pearson correlation coefficient for different subsets of and ,in express and Only include the same target or The change in the causal effect when express and Also includes the target and , in order to further test the combinatorial reasoning ability of the visual encoder.

[0061] Step 3: Explain and evaluate the synergistic effect of the visual encoder and the text encoder: Based on the explanation and evaluation results of steps 1 and 2, further measure the correlation between the specific interaction relationship of the image entity and the specific combination relationship of the text entity, so as to quantitatively evaluate the synergistic effect of the visual encoder and the text encoder; specifically, the following sub-steps are included:

[0062] 3.1. Analyze the impact of text changes on the collaborative output of the visual language model: Based on the sensitivity evaluation function in step 1.2 , measuring the influence of specific components of the image on the interaction between text object words and relation words , to explain when only given When the image area of ​​the neutron object is and How changes in the interaction between object words and relation words within a sentence will affect the output of the visual language model.

[0063] 3.2. Analyze the impact of image changes on the collaborative output of the visual language model: According to the sensitivity evaluation function in step 1.2 , measuring the influence of specific components of the image on the interaction between text object words and relation words :

[0064]

[0065] in, is the normalization coefficient, for In terms of Represents object words and relation words in the text. By measuring the correlation between specific components of the image and the interaction between object words and relation words, we can get and , and use it to diagnose the visual language model. Measured when only given When the image area of ​​the neutron object is and The degree to which the interaction between object words and relation words in the visual language model affects the output.

[0066] 3.2. According to the sensitivity evaluation function in step 2.2 , measuring the influence of specific components of text on the interaction between image objects :

[0067]

[0068] in is the normalization coefficient, for In terms of Representation Object and By measuring the correlation between the text component and the interaction between the objects on the image, we can get and , and use this to check the visual language model. Measured when only given When the object word in and The degree to which changes in the relationship between objects within a visual language model affect the output.

[0069] An embodiment of the present invention also provides an electronic device, comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement a robustness diagnosis method for fine-grained perception capabilities of a large visual language model as described in any one of the embodiments.

[0070] An embodiment of the present invention also provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, a robustness diagnosis method for fine-grained perception capability of a large visual language model in the above embodiment is implemented.

[0071] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capability and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0072] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A robust diagnostic method for fine-grained perception capabilities of large visual language models, It is characterized in that The steps include: (1) Diagnose semantic interaction relationships from the perspective of language entities and evaluate the text encoder's ability to understand the changing relationships between text patterns; (2) Evaluate the ability to perceive fine-grained differences in the visual object perspective and evaluate the ability of the visual encoder to understand the changing relationship between different visual targets; (3) Combining the evaluation results of the collaboration between the language entity perspective and the visual object perspective to evaluate the understanding capabilities of the visual encoder and the text encoder for the collaborative object attribute combination relationship.

2. According to claim 1, a robustness diagnosis method for fine-grained perception capability of a large visual language model, It is characterized in that The step (1) is specifically as follows: on the premise of keeping the text entities unchanged but changing the combination mode of the text entities, the Hassani dividend metric is used to quantitatively evaluate the sensitivity of the text encoder to the change of the text mode, and to evaluate the text encoder's ability to understand the combination relationship between the text entities.

3. According to claim 2, a robustness diagnosis method for fine-grained perception capability of a large visual language model, It is characterized in that The step (1) includes the following sub-steps: (1.1) Given a set of image-text pairs , by exchanging Get the perturbation text from the entity words in ; Align image text and Input the visual language model separately to obtain two matching scores: and ;in, represents the original image, Represents the original text, represents the perturbed text obtained by exchanging entity words, Represents all visual concepts on the original image, represents all the words in the original text, represents all the words in the perturbation text; (1.2) Using the matching score obtained in step (1.1) and , in a given image and words , Based on different subsets of Quantitatively measure the sensitivity of text encoders to text patterns; represents a subset of words in the original text, represents a subset of words in the perturbed text; (1.3) Using the matching score obtained in step (1.1) and , calculate the reward difference and interaction effect differences , and then calculate the word and Pearson correlation coefficient for different subsets of , and , to test the combinatorial reasoning ability of the text encoder; among them, Indicates when and The difference in the interaction effect only includes the object words, Indicates when and The difference in the interaction effect only includes the relational terms, Indicates when and The differences in the interactive effects of both relation words and object words are included.

4. According to claim 3, a robustness diagnosis method for fine-grained perception capability of a large visual language model, It is characterized in that In the step (1.2), the sensitivity assessment includes: Sensitivity assessment, attribute words Sensitivity assessment, target words Sensitivity assessment, relational words and object words Sensitivity assessment of interactions and attribute and object terms Sensitivity assessment of interactions.

5. According to claim 1, a robustness diagnosis method for fine-grained perception capability of a large visual language model, It is characterized in that The step (2) specifically comprises: given two images containing the same objects but with significantly different relationships between the objects; quantifying the change in the causal effect of the visual encoder output, analyzing the sensitivity of the visual encoder to the change in the relationship between different object components; and evaluating the visual encoder's ability to understand the combined relationship between visual targets.

6. According to claim 5, a robustness diagnosis method for fine-grained perception capability of a large visual language model, It is characterized in that The step (2) includes the following sub-steps: (2.1) Given a set of image-text pairs , by changing The entity interaction relationship in obtains the perturbed image ; Align image text and Input the visual language model separately to obtain two matching scores: and ;in, represents the original image, Indicates changing the entity interaction relationship of the original image to obtain the perturbed image. Represents the original text, Represents all visual concepts on the original image, represents all the words in the original text, Represent all visual concepts on the perturbed image; (2.2) Using the matching score obtained in step (2.1) and , in a given text and visual concepts and Based on different subsets of To quantitatively analyze the sensitivity of the visual encoder to two types of fine-grained differences between image objects: Sensitivity evaluation and paired objects The sensitivity assessment of the relationship between them is used to diagnose the visual encoder's ability to understand visual components; among them, A subset of visual concepts representing the original image, A subset of visual concepts representing the perturbed image, express and Each has its own object area, express and Two object areas at the same time; (2.3) Using the matching score obtained in step (2.1) and , calculate the reward difference and interaction effect differences , and then calculate the visual concept and Pearson correlation coefficient for different subsets of and , to test the combinatorial reasoning ability of the visual encoder; among them, express and The difference in the interaction effect of the object areas they each have, express and The difference in the interactive effect of two object regions simultaneously.

7. According to claim 1, a robustness diagnosis method for fine-grained perception capability of a large visual language model, It is characterized in that The step (3) is specifically as follows: based on the interpretation and evaluation results of the steps (1) and (2), further measuring the correlation between the specific interaction relationship of the image entities and the specific combination relationship of the text entities, so as to quantitatively evaluate the synergistic effect of the visual encoder and the text encoder.

8. According to claim 7, a robustness diagnosis method for fine-grained perception capability of a large visual language model, It is characterized in that The step (3) includes the following sub-steps: (3.1) Based on the sensitivity evaluation function , measuring the influence of specific components of the image on the interaction between text object words and relation words , to explain when only given When the image area of ​​the neutron object is and How will the interaction between object words and relation words in the text affect the output of the visual language model? represents the original image, Represents the original text, represents the perturbed text obtained by exchanging entity words, Indicates a fixed original image Sub-image area, change the original text object words and relation words; (3.2) Based on the sensitivity evaluation function , measuring the influence of specific components of text on the interaction between image objects , to explain when only given When the object word in and How changes in the relationship between objects in the visual language model affect the output; represents the original image, Indicates changing the entity interaction relationship of the original image to obtain the perturbed image. Represents the original text, Indicates fixed original text The object word changes the original image object.

9. An electronic device comprising a memory and a processor, It is characterized in that The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement a robust diagnosis method for fine-grained perception capabilities of a large visual language model as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the program is executed by a processor, it implements a robustness diagnosis method for fine-grained perception capabilities of a large visual language model as described in any one of claims 1-8.

Citation Information

Patent Citations

  • Image pair difference description method based on coding-decoding end

    CN114581690A