A multi-modal large language model neuron attribution method and related device

Abstract

The application discloses a multimodal large language model neuron attribution method and related equipment, the method comprises the following steps: generating an image with specific semantics by using an MLMs model, and obtaining a specific region related to the specific semantics in the image; evaluating the correlation between each dimension input into the image generation module of the MLMs model and the specific region in the generated image; obtaining the contribution of neurons to the last representation, integrating the contribution scores of each dimension, and taking the neuron with the highest contribution score as an I-neuron; obtaining an attribution score of neurons to text output, and taking the neuron with the highest attribution score as a T-neuron. The application does not depend on time-consuming and extremely high storage requirement technologies such as gradient calculation and causal effect, can exclude the interference of semantic noise information in the generated image, and finds key neurons for different output modalities by decoupling for specific semantic concepts, so that efficient and accurate neuron attribution is realized.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, specifically a neuron attribution method and related equipment for multimodal large language models. Background Technology

[0002] As Large Language Models (LLMs) demonstrate remarkable capabilities, revealing their internal mechanisms becomes increasingly important, especially in applications emphasizing fairness, trust, and ethical decision-making. "Neural attribution" is a key approach to interpreting LLMs. This method reveals the knowledge and linguistic properties learned by individual model components, such as neurons and hidden layers, by attributing text output to them. These insights not only aid in model editing and parameter optimization but also deepen our understanding of how LLMs recognize and store diverse input information. Current neuronal attribution research suggests that this storage capacity may primarily stem from their feedforward neural networks (FFNs).

[0003] Recently, the rapid development of multimodal large-scale language models (MLLMs) has sparked considerable interest among researchers in their interpretability, especially MLLMs capable of simultaneously perceiving and generating text and images. These MLLMs typically include two additional common components on top of basic text LLMs: an image encoding module and a generation module. Specifically, the image encoding module projects the input image into the representation space of the base LLMs; subsequently, the image generation module generates the image output based on the representations provided by the base LLMs. Taking the classic large image generation model (GILL) as an example, such as... Figure 1 As shown, it uses OPT as the base LLMs, CLIP Vit-L with a mapper as the image encoding module, and StableDiffusion Model and another mapper as the image generation module.

[0004] While this scalability enables GILL to perform a variety of downstream tasks, it also presents challenges to its interpretation, particularly regarding neuronal attribution. Specifically, we outline these challenges through three progressively deeper points:

[0005] First, the challenge of attribution source: semantic noise. For example... Figure 2 As shown in (a), Figure 2Figure (a) illustrates the neuronal attribution process of LLMs that only output text. In the figure, for semantics such as "dog," current methods typically attribute the output word "dog" directly to neurons in the LLMs. However, in MLLMs, directly attributing the entire generated image to neurons can lead to inaccuracies. Figure 2 As shown in (b), Figure 2 (b) shows the output attribution noise map of MLLMs, which is a transfer of the neuronal attribution process. When GILL is asked to draw a dog, the generated image may contain other semantic elements, such as a lawn, which introduces noise and interferes with the attribution process.

[0006] Second, the attribution process itself presents a challenge: inefficiency. Current attribution methods typically rely on gradients or causal effects, both of which require extensive forward and backward propagation processes, which are extremely time-consuming and memory-intensive. Furthermore, the additional complexity introduced by the encoding and generation modules in MLLMs further exacerbates this challenge.

[0007] Third, the challenge of attribution results: neuronal hybridity. In LLMs that only output text, attributing the concept of "dog" involves identifying neurons crucial to the output of the word "dog," called T-neurons. In contrast, MLLMs also require identifying neurons crucial to image generation, called I-neurons. Figure 2 As shown in (b), T-neurons and I-neurons can have different distributions even for the same concept, which can lead to complex and conflicting attribution results, making further analysis and application more difficult.

[0008] In summary, current research on MLLMs urgently needs to find a neuronal attribution paradigm that can efficiently and accurately attribute the output of MLLMs. Summary of the Invention

[0009] This embodiment provides a multimodal large language model neuron attribution method and related equipment to solve the problems of semantic noise, inefficiency and neuron hybridity in the attribution process in related technologies.

[0010] In a first aspect, the present invention provides a multimodal large language model neuron attribution method, the multimodal large language model neuron attribution method comprising:

[0011] Using MLLMs models to generate images with specific semantics, specific regions related to specific semantics are obtained from the images;

[0012] The correlation between each dimension of the input to the image generation module of the MLLMs model and a specific region in the generated image is evaluated, and the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained based on the correlation.

[0013] Obtain the contribution of each neuron to the final representation, integrate the contribution scores of each dimension, and designate the neuron with the highest contribution score as the I-neuron; obtain the attribution score of each neuron for the text output, and designate the neuron with the highest attribution score as the T-neuron.

[0014] In an optional embodiment, obtaining specific regions in an image with specific semantics generated using MLLMs models includes:

[0015] Input text information with specific semantics into the MLLMs model;

[0016] An image with specific semantics is generated by the MLLMs model based on the input text information;

[0017] A pre-trained segmentation model is used to identify specific regions in the image that are related to a specific semantic meaning.

[0018] In an optional embodiment, the correlation between each dimension of the input to the MLLMs model image generation module and a specific region in the generated image is evaluated. The contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained based on the correlation, including:

[0019] The correlation score is obtained by evaluating the correlation between each dimension of the input to the image generation module of the MLLMs model and a specific region in the generated image using the diffusion model attribution algorithm.

[0020] The correlation scores are normalized to obtain the contribution scores of each element in the final representation of the Lth layer of the MLLMs model relative to the semantics of the image target.

[0021] In an optional embodiment, the contribution of each neuron to the final representation is obtained, and the contribution scores for each dimension are integrated. The neuron with the highest contribution score is selected as the I-neuron.

[0022] Obtain the direct and indirect contributions of neurons to the final representation.

[0023] In an optional embodiment, obtaining the direct contribution of a neuron to the final representation includes:

[0024] Based on neuron activation, the process of generating the final representation is obtained;

[0025] Based on the generation process of the final representation, the direct contribution of the neuron to the final representation is obtained.

[0026] In an optional embodiment, obtaining the indirect contribution of a neuron to the final representation includes:

[0027] Design a maximum value function for the contribution score:

[0028]

[0029] in, Indicates neuronal contribution. This indicates neuronal activation.

[0030] In an optional embodiment, obtaining the attribution score of the neuron for the text output, and designating the neuron with the highest attribution score as the T-neuron, includes:

[0031] Obtain the contribution of residual flow to the output of MLLMs models;

[0032] The contribution of individual neurons is refined, and the neurons that contribute the most to a single semantic concept are retained;

[0033] The final attribution score of the neuron for the text output is obtained, and the neuron with the highest attribution score is designated as the T-neuron.

[0034] Compared with existing technologies, the beneficial effects of the multimodal large language model neuron attribution method of the present invention are as follows:

[0035] This invention proposes a neuron attribution method for multimodal large language models that does not rely on time-consuming and storage-intensive techniques such as gradient calculation and causal effects. It also eliminates semantic noise interference inherent in generated images. For specific semantic concepts, it decouples and identifies key neurons (T neurons and I neurons) for different output modalities, achieving efficient and accurate neuron attribution. This method can be used to attribute and explain the predictive behavior of multimodal large language models, while extracting interpretable neurons with class-aware semantic capabilities from MLLMs. It overcomes the inherent computational time and noise introduction problems of traditional attribution methods, eliminates the influence of interfering semantics irrelevant to the target semantics, and effectively improves the accuracy and reliability of the attribution method.

[0036] Secondly, the present invention provides a multimodal large language model neuron attribution device, comprising:

[0037] The image generation module is used to obtain specific regions in the image that are related to the specific semantics, based on images with specific semantics generated by MLLMs models.

[0038] The contribution score calculation module is used to evaluate the correlation between each dimension of the input of the MLLMs model image generation module and a specific region in the generated image. Based on the correlation, the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained.

[0039] The attribution module is used to obtain the contribution of neurons to the final representation, integrate the contribution scores of each dimension, and take the neuron with the highest contribution score as the I-neuron; obtain the attribution score of neurons for text output, and take the neuron with the highest attribution score as the T-neuron.

[0040] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multimodal large language model neuron attribution method as described in the first aspect.

[0041] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the multimodal large language model neuron attribution method as described in the first aspect.

[0042] Compared with the prior art, the beneficial effects of the multimodal large language model neuron attribution device, electronic device and storage medium of the present invention are the same as those of the multimodal large language model neuron attribution method described in the first aspect, and therefore will not be repeated here. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A schematic diagram of the general paradigm for existing MLLMs;

[0045] Figure 2 This is a schematic diagram of the semantic noise interference attribution process;

[0046] Figure 3 This is a network architecture diagram of base LLMs in a multimodal large language model in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of the neuronal attribution paradigm for image output in an embodiment of the present invention;

[0048] Figure 5This is a schematic diagram of the neuronal attribution paradigm for text output in an embodiment of the present invention;

[0049] Figure 6 This is a distribution map of T / I neurons identified by ATN in an embodiment of the present invention;

[0050] Figure 7 This is a schematic diagram illustrating the average invariance of different concepts in embodiments of the present invention;

[0051] Figure 8 This is a schematic diagram illustrating the average semantic specificity of different concepts in an embodiment of the present invention;

[0052] Figure 9 This is a flowchart of the multimodal large language model neuron attribution method in an embodiment of the present invention;

[0053] Figure 10 This is a structural block diagram of the multimodal large language model neuron attribution device in an embodiment of the present invention;

[0054] Figure 11 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0055] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.

[0056] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0057] This invention provides a neuron attribution method for multimodal large language models (ATN). Figure 9 This is a flowchart of the multimodal large language model neuron attribution method of the present invention, as shown below. Figure 9 As shown, the process includes the following steps:

[0058] S100. Using MLLMs models to generate images with specific semantics, obtain specific regions in the images that are related to specific semantics;

[0059] It should be noted that MLLMs, or Multimodal Large Language Models, are based on Large Language Models (LLMs). They add an image encoding module and a generation module to the LLMs model. Using MLLMs, images with semantic correspondence to the input text information can be generated.

[0060] like Figure 3 As shown, Figure 3 This is a network architecture diagram of base LLMs in a multimodal large language model. First, let's explain the network architecture of base LLMs in a multimodal large language model:

[0061] Its large language model based on autoregressive Transformer Based on vocabulary It performs the operation. Specifically, it receives an input sequence of tags. And generate probability distribution To predict the next tag. Each tag is in the first... Layers are represented as a series of representations ,in, It is a marker The embedding. The final output of the model. From the The last representation of the layer Using the anti-embedding matrix Obtain. For example... Figure 3 As shown, Figure 3 It shows In the The computation process within a layer. The representation of each layer is determined by global attention. Local MLP output and the previous layer's representation It is composed of various elements. The formula is expressed as:

[0062]

[0063] in, These are the first and second linear layers in an FFN, with the intermediate layer having a dimension of . ; and These are rectification and normalization nonlinear operations. Furthermore, for the... The first layer of FFN one neuron Consider activating neurons The definition can be simplified to:

[0064] ;

[0065] Based on the internal architecture of the aforementioned base LLMs, common MLLM paradigms include: Figure 3 As shown. Specifically, MLLMs have made the following modifications based on the LLMs mentioned above:

[0066] In addition to text prompts, input a marked sequence. It also includes the encoding of the input image generated by the image encoding module;

[0067] The representation of the last hidden layer is used as input to the image generation module, facilitating conditional image generation.

[0068] Based on the above, using MLLMs to generate images of specific concepts (e.g., dogs) often produces outputs containing redundant noise elements (e.g., lawns). To mitigate the negative impact of semantic noise on neuron attribution, a two-step method is proposed to extract I-Neurons. The first step focuses on attributing the output of the image generation module (i.e., the image) to the input of the image generation module. The second step focuses on inputting the image generation module. This is attributed to specific neurons.

[0069] Specifically, using MLLMs models to generate images with specific semantics, obtaining specific regions in the image that are related to specific semantics includes:

[0070] Input text information with specific semantics into MLLMs models; for example, text information with the semantics of "dog";

[0071] Based on the input text information, images with specific semantics are generated by MLLMs models; MLLMs models generate images containing the semantics of "dog".

[0072] Pre-trained segmentation models are used to identify specific regions in an image that are related to specific semantics.

[0073] Use segmentation models (such as EVA02) to identify specific regions associated with "dog" in order to minimize interference from redundant semantics in the background (such as lawn).

[0074] S200. Evaluate the correlation between each dimension of the input to the MLLMs model image generation module and a specific region in the generated image, and obtain the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image based on the correlation.

[0075] The correlation between each dimension of the input to the image generation module of the MLLMs model and a specific region in the generated image is evaluated. Based on the correlation, the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained, including:

[0076] The correlation score is obtained by evaluating the correlation between each dimension of the input to the image generation module of the MLLMs model and a specific region in the generated image using the diffusion model attribution algorithm.

[0077] For example, the Diffuser-Interpreter algorithm is used to evaluate the correlation between each dimension of the input to the image generation module of the MLLMs model and the "dog" region in the generated image; a correlation score is obtained.

[0078] The correlation scores are normalized to obtain the contribution scores of each element in the final representation of the Lth layer of the MLLMs model relative to the semantics of the image target.

[0079] Specifically, the final representation is obtained by normalizing these correlation scores to (0, 1). The contribution score (importance score) of each element relative to the target semantics of the output image. .

[0080] It should be noted that, by using semantic segmentation algorithms and attribution algorithms for generative modules tailored for other modalities (such as audio and video), transfer to any other modality can be achieved. This is achieved by acquiring [the necessary information] through these modality-specific algorithms. Subsequently, the attribution steps are common to all Transformer-based MLLMs.

[0081] The purpose of the above steps is to attribute the image to the final representation. That is, to identify the final representation The contribution of each element to image generation.

[0082] S300: Obtain the contribution of each neuron to the final representation, integrate the contribution scores of each dimension, and take the neuron with the highest contribution score as the I-neuron; obtain the attribution score of each neuron to the text output, and take the neuron with the highest attribution score as the T-neuron.

[0083] Obtain the contribution of each neuron to the final representation, integrate the contribution scores for each dimension, and select the neuron with the highest contribution score as the I-neuron.

[0084] Obtain the direct and indirect contributions of neurons to the final representation.

[0085] It should be noted that current contribution scoring methods typically rely on gradients, such as the product of gradient and activation or the integral of gradient. However, these methods are computationally expensive for large-scale models, especially those with many parameters. Therefore, inspired by research on neuronal activation, we propose a novel attribution scoring method that relies solely on neuronal activation. .

[0086] like Figure 4 As shown, specifically, obtaining the direct contribution of a neuron to the final representation includes:

[0087] Based on neuron activation, the process of generating the final representation is obtained;

[0088] Based on the generation process of the final representation, the direct contribution of the neuron to the final representation is obtained.

[0089] First, the final representation is expanded using the following formula. To decompose and derive the final representation The generation process:

[0090]

[0091] in It is a weight matrix The Columns, corresponding to neurons The index. The first term of the last row of the above formula reflects the neuron. The last statement The direct contribution is the contribution through the residual flow of the underlying LLMs. Therefore, the following is adopted: As neurons For the final representation Indicators of contribution.

[0092] In addition, by integrating the importance scores of each dimension This allows for the establishment of a complete attribution paradigm (i.e., generating the target semantic region in the image → representing...). →neuron Importance score This is already the final statement. Each dimension in the neuron is assigned a contribution. Contribution It can be defined as:

[0093] .

[0094] Furthermore, obtaining the indirect contribution of neurons to the final representation includes:

[0095] Design a maximum value function for the contribution score:

[0096]

[0097] in, Indicates neuronal contribution. This indicates neuronal activation.

[0098] It should be noted that, although neurons contribute Quantified neurons The final representation is obtained through residual flow. It considers the direct contribution of the current index neuron, but it doesn't take all the influencing factors into account. Specifically, it ignores the indirect contributions made by neurons through the attention mechanism within subsequent FFN layers. To address this issue and avoid complex calculations like gradients, the contribution of the current index neuron is further evaluated. Optimize.

[0099] Furthermore, the relative magnitude of neuronal activation is used as another indicator to identify potential factors affecting the final representation. Neurons that make significant indirect contributions, these contributions may be made by neurons. neglect.

[0100] Furthermore, considering the contribution of neurons The calculation already includes neuron activation. The contributions reflected by these two metrics may overlap. To avoid redundancy caused by adding or multiplying these two metrics, the aforementioned maximum value function was used in the final scoring design.

[0101] Among these, the normalization operation ensures the neuron's contribution. and neuronal activation Fair competition between them, and their complementarity, make this combined scoring method superior to using neuron contributions alone. and neuronal activation More effective. By calculating the contribution of different semantics at each level, this method identifies the neurons that consistently exhibit the highest contribution in the generated image. These neurons are designated as I-neurons responsible for specific semantics in MLLMs due to their significant role.

[0102] Furthermore, the attribution scores of neurons for the text output are obtained, and the neurons with the highest attribution scores are designated as T-neurons, including:

[0103] Obtain the contribution of residual flow to the output of MLLMs models;

[0104] The contribution of individual neurons is refined, and the neurons that contribute the most to a single semantic concept are retained;

[0105] The final attribution score of the neuron for the text output is obtained, and the neuron with the highest attribution score is designated as the T-neuron.

[0106] Similar to the derivation process of image output contribution scoring, we first focus on the impact of residual flow on the output. The contribution. Specifically, will Expanded as follows:

[0107]

[0108]

[0109] Based on common assumptions used to analyze the internal mechanisms of LLMs, the representation of any layer in a language model can be transformed into a distribution over a labeled vocabulary using output embeddings. Therefore, It can be considered as a new anti-embedding matrix at the end of the residual flow. pass Model output distribution Contributions to the vocabulary.

[0110] Furthermore, such as Figure 5 As shown, this contribution is refined to specific neurons to further decompose the above equations and refine the individual neuron level. Contribution to the output word. For example, let... Index for "dog" in the vocabulary:

[0111]

[0112]

[0113] in yes The Okay. According to the above equation, the neuron For the vocabulary list The contribution of the word "dog" for:

[0114]

[0115] Furthermore, the identified key neurons exhibit semantic specificity (i.e., only neurons that contribute the most to a single semantic concept are retained, while other semantics are ignored). Formally, for the first term in the vocabulary... The semantics of a word, neurons Contribution It can be defined as:

[0116] for

[0117] Based on this, neurons can be obtained. The final attribution score for the text output is used to designate the neurons that consistently show the highest contribution as T-neurons.

[0118] For example, based on the distribution of T / I neurons identified by ATN, 1000 images are randomly selected from the COCO dataset. Each image is then input into GILL, instructing GILL to generate similar images. The distribution of T / I neurons is as follows: Figure 6 As shown.

[0119] from Figure 6As can be seen, in MLLMs, key neurons containing semantically specific text and image outputs are mainly located in the mid-to-high levels of the basic LLMs. This observation is consistent with previous studies on the attribution of neurons within LLMs. Furthermore, the similar distribution of T neurons and I neurons suggests that the semantic concept formation time of different modalities may be consistent in MLLMs.

[0120] Figure 6 This study revealed partial overlap between T neurons and I neurons. However, it also clearly demonstrated the differences between them. This finding confirms that even with the same semantics, key neurons are modality-specific in MLLMs.

[0121] For different text / image outputs containing the same semantics, the T / I neurons identified by the attribution method should be consistent. To quantify this consistency, GILL is instructed to describe and generate images with the same semantics ten times, collecting a set of T / I neurons each time. The proportion of neurons appearing in all ten sets is then calculated to quantify cross-sample invariance. Figure 7 It demonstrates the average invariance of different concepts.

[0122] according to Figure 7 The following conclusions can be drawn: ATN outperforms all baselines by an average of 16.83% in terms of cross-sample invariance across all semantics. This indicates that ATN is able to extract key neurons for the target semantics and effectively filter out neurons that are sensitive to sample-specific noise.

[0123] Neurons that are crucial for a specific semantic meaning should not be indiscriminately important in other semantic meanings. Therefore, this section studies the specificity of neurons and identifies the top 500 T / I neurons for a specific semantic meaning. Then, it shows the number of neurons that are important only for that specific semantic meaning, and the number of neurons with a probability lower than [a certain threshold]. In this case, the average number of neurons important for other semantics is shown. The results of the three best-performing methods are presented here. Figure 8 As shown.

[0124] according to Figure 8 The following conclusions can be drawn: ATN outperforms all baselines by an average of 19.94% in cross-sample invariance across all semantics, demonstrating that the T / I neurons identified by ATN are specialized and not universally sensitive across different semantics, thus verifying their semantic specificity.

[0125] In summary, this invention can provide attribution and explanation for various applications in multimodal large language models, while extracting interpretable neurons with class-aware semantic capabilities from MLLMs. This invention overcomes the inherent computational time and noise introduction problems of traditional attribution and explanation methods, eliminates the influence of interfering semantics irrelevant to the target semantics, and effectively improves the accuracy and reliability of attribution methods.

[0126] This invention also provides a multimodal large language model neuron attribution device for implementing the above-described method embodiments; details already described will not be repeated. The terms "module," "unit," and "subunit," etc., used below refer to combinations of software and / or hardware that perform a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation or a combination of software and hardware is also possible and contemplated.

[0127] like Figure 10 As shown, Figure 10 This is a structural block diagram of the multimodal large language model neuron attribution device of the present invention, which includes:

[0128] The image generation module is used to obtain specific regions in the image that are related to the specific semantics, based on images with specific semantics generated by MLLMs models.

[0129] The contribution score calculation module is used to evaluate the correlation between each dimension of the input of the MLLMs model image generation module and a specific region in the generated image. Based on the correlation, the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained.

[0130] The attribution module is used to obtain the contribution of neurons to the final representation, integrate the contribution scores of each dimension, and take the neuron with the highest contribution score as the I-neuron; obtain the attribution score of neurons for text output, and take the neuron with the highest attribution score as the T-neuron.

[0131] Figure 11 A structural block diagram of the electronic device provided in the embodiments of the present invention, such as... Figure 11 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the following methods:

[0132] Using MLLMs models to generate images with specific semantics, specific regions related to specific semantics are obtained from the images;

[0133] The correlation between each dimension of the input to the image generation module of the MLLMs model and a specific region in the generated image is evaluated, and the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained based on the correlation.

[0134] Obtain the contribution of each neuron to the final representation, integrate the contribution scores of each dimension, and designate the neuron with the highest contribution score as the I-neuron; obtain the attribution score of each neuron for the text output, and designate the neuron with the highest attribution score as the T-neuron.

[0135] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0137] Using MLLMs models to generate images with specific semantics, specific regions related to specific semantics are obtained from the images;

[0138] The correlation between each dimension of the input to the image generation module of the MLLMs model and a specific region in the generated image is evaluated, and the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained based on the correlation.

[0139] Obtain the contribution of each neuron to the final representation, integrate the contribution scores of each dimension, and designate the neuron with the highest contribution score as the I-neuron; obtain the attribution score of each neuron for the text output, and designate the neuron with the highest attribution score as the T-neuron.

[0140] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A neuronal attribution method for a multimodal large language model, characterized in that, The multimodal large language model neuron attribution method includes: The MLLMs model is used to generate an image corresponding to a specific semantic in the input text, and the image region corresponding to the specific semantic in the image is obtained by the segmentation model. The correlation between each dimension of the input to the image generation module of the MLLMs model and the image region corresponding to a specific semantic in the generated image is evaluated, and the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained based on the correlation. Obtain the contribution of each neuron to the final representation, integrate the contribution scores of each dimension, and designate the neuron with the highest contribution score as the I-neuron; obtain the attribution score of each neuron for the text output, and designate the neuron with the highest attribution score as the T-neuron.

2. The multimodal large language model neuron attribution method according to claim 1, characterized in that, The correlation between each dimension of the input to the image generation module of the MLLMs model and the image region corresponding to a specific semantic in the generated image is evaluated. Based on the correlation, the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained, including: The correlation score is obtained by evaluating the correlation between each dimension of the input to the image generation module of the MLLMs model and the image region corresponding to a specific semantic in the generated image using the diffusion model attribution algorithm. The correlation scores are normalized to obtain the contribution scores of each element in the final representation of the Lth layer of the MLLMs model relative to the semantics of the image target.

3. The multimodal large language model neuron attribution method according to claim 1, characterized in that, Obtain the contribution of each neuron to the final representation, integrate the contribution scores for each dimension, and select the neuron with the highest contribution score as the I-neuron. Obtain the direct and indirect contributions of neurons to the final representation.

4. The multimodal large language model neuron attribution method according to claim 3, characterized in that, Obtaining the direct contribution of neurons to the final representation includes: Based on neuron activation, the process of generating the final representation is obtained; Based on the generation process of the final representation, the direct contribution of the neuron to the final representation is obtained.

5. The multimodal large language model neuron attribution method according to claim 3, characterized in that, Obtaining the indirect contribution of neurons to the final representation includes: Design a maximum value function for the contribution score: ； in, Indicates neuronal contribution. This indicates neuronal activation.

6. The multimodal large language model neuron attribution method according to claim 1, characterized in that, Obtain the attribution scores of neurons for the text output, and select the neuron with the highest attribution score as a T-neuron. Obtain the contribution of residual flow to the output of MLLMs models; The contribution of individual neurons is refined, and the neurons that contribute the most to a single semantic concept are retained; The final attribution score of the neuron for the text output is obtained, and the neuron with the highest attribution score is designated as the T-neuron.

7. A multimodal large language model neuronal attribution device, characterized in that, include: The image generation module is used to generate an image corresponding to a specific semantic in the input text using MLLMs models, and to obtain the image region corresponding to the specific semantic in the image through a segmentation model. The contribution score calculation module is used to evaluate the correlation between each dimension of the input of the MLLMs model image generation module and the image region corresponding to a specific semantic in the generated image. Based on the correlation, the contribution score of each element in the final representation of the Lth layer of the MLLMs model relative to the target semantics of the image is obtained. The attribution module is used to obtain the contribution of neurons to the final representation, integrate the contribution scores of each dimension, and take the neuron with the highest contribution score as the I-neuron; obtain the attribution score of neurons for text output, and take the neuron with the highest attribution score as the T-neuron.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multimodal large language model neuron attribution method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the multimodal large language model neuron attribution method as described in any one of claims 1 to 6.

Images