123 results about "Information bottleneck method" patented technology

The invention discloses a multi-modal multi-scale retrieval enhancement generation method, system and device applied to external knowledge questions and answers and a medium. The method comprises question perception, multi-modal multi-scale query fusion coding, dense recall and answer generation. Analyzing a key query phrase from the question through a fine-tuned instruction language model, and accurately positioning a region of interest corresponding to the phrase in an image by using an open set visual positioning model; multi-source information is compressed and distilled into an optimal query vector through a deep fusion network integrating multi-head self-attention and an information bottleneck theory; executing a maximum inner product search to recall related knowledge; guiding the large language model to synthesize all information to generate a final answer; the system, the equipment and the medium directly perform feature fusion in the vector space based on the method, so that challenges such as information loss and cascading errors caused by a traditional normal form can be effectively dealt with, high correlation and high accuracy of retrieval knowledge are ensured, and accurate and reliable image-text questions and answers are realized.

The invention discloses a source load risk management and control method for cooperative operation of multiple types of adjusting power supplies and new energy, and the method comprises the steps: depicting a prediction error tail dependency relationship through employing an extreme value theory and a Copula function, constructing a distribution robust optimization model in combination with a Wasserstein distance, and obtaining the standby capacity sharing, triggering threshold and risk index of the multiple types of adjusting power supplies; performing conformal prediction based on the prediction residual error to generate a dynamic confidence interval, and mapping the dynamic confidence interval into a pipeline radius of model prediction control; executing rolling MPC containing information bottleneck regularization under the constraint of the reserve capacity and the pipeline, and generating a robust scheduling strategy; and constructing a polyhedral feasible region model of the flexible load group according to the scheduling strategy and the risk index, and performing scaling to obtain a group control action. And the new energy consumption level and the system operation safety are obviously improved.

The invention discloses a data analysis and prediction method and system based on multi-modal data fusion, and the method comprises the steps: collecting multi-modal original data, and carrying out the preprocessing of the multi-modal original data, and generating normalized data; setting sliding window parameters, and performing sliding fragmentation on each modal data to obtain sliding window fragment data; extracting modal features, and fusing the modal features into sliding window level fusion feature representation; inputting the improved variational information bottleneck model, and performing feature compression and reconstruction; all fusion features are aggregated, a prediction result is output through modeling reasoning, and the model is optimized. According to the method, the improved sliding window algorithm and the variational information bottleneck model are fused, so that efficient analysis and accurate prediction of multi-modal data are realized.

The invention relates to the technical field of knowledge maps, in particular to a few-sample knowledge map completion method and system fusing relation perception information bottleneck. The method comprises the following steps: S1, preprocessing an input triple; s2, building a global aggregation module, and updating entity embedding; s3, establishing a relationship aggregation module, and updating relationship embedding; s4, establishing a relationship-based information bottleneck module, filtering noise irrelevant to tasks, and meanwhile, retaining relationship-specific information; s5, building an EM attention pooling module, adaptively aggregating multi-path semantic representation, and highlighting the correlation between the entity and the relationship; and S6, establishing a score calculation module, calculating a triple score and outputting the triple score. The invention provides a few-sample knowledge graph completion method and a few-sample knowledge graph completion system fusing relation perception information bottleneck, which are used for solving the problems of insufficient relation and entity representation coupling, high redundant information interference and difficulty in modeling due to high-order relation dependence in a knowledge graph completion task, and realizing efficient inference of potential relation facts.

The invention discloses a single-classification industrial control system anomaly detection method based on double-view information bottleneck fusion, which belongs to the technical field of industrial control system anomaly detection and comprises the following steps of: in a training stage, simultaneously constructing a time sequence view and a pseudo image view; respectively extracting characterization through a time dynamic encoder and a pre-trained visual backbone; performing cross-view compression in the variational information bottleneck fusion module and establishing projection branches only depending on sequence representation; a normal subspace is learned around a center vector, an accurate hypersphere is established to describe normal data, and a center distance is used as an abnormal score and a quantile threshold is used to complete discrimination; in the detection stage, online recognition can be performed only by inputting a time sequence view. According to the method, dual-view prior and variational information bottleneck are fused, end-to-end collaborative optimization is carried out on the dual-view prior and the variational information bottleneck and deep support vector data description, an accurate decision boundary can be established under the condition of a small number of samples, and a solution for industrial control system anomaly detection under the condition that the samples are limited is provided.

The invention belongs to the field of data detection, and provides a false news detection method and system based on fact-emotion dual uncertainty, and the method comprises the steps: obtaining a text and an image of each piece of news in a social network; the method comprises the following steps: performing feature extraction on a text and an image to obtain text embedding and image embedding, performing Gaussian reweighting on the text embedding and the image embedding respectively, then performing calculation to generate Gaussian uncertainty representations, and screening the Gaussian uncertainty representations based on a variational information bottleneck strategy to obtain multi-modal uncertainty representations; processing the multi-modal uncertainty representation to generate a fact inconsistent representation, and processing the multi-modal uncertainty representation based on an emotion manipulation graph convolutional network to generate an emotion inconsistent representation; and fusing the fact inconsistent representation and the emotion inconsistent representation to generate fusion features, and performing classification based on the fusion features to obtain a false news detection classification result. According to the method, the limitation of cross-modal complementarity caused by the uncertainty of multi-modal data is relieved.

The invention relates to the technical field of data processing, in particular to a large model reasoning optimization method based on context incremental updating, which comprises the following steps: processing a multi-modal data stream through timestamp alignment and a filtering algorithm, extracting features by adopting a shared encoder and a private encoder, and realizing feature decoupling through a depth information bottleneck principle. A dynamic emotion map is constructed by using Gaussian process regression and a random process algorithm, and self-adaptive updating control is realized by combining meta-learning and Bayesian optimization. Incremental state management is realized by adopting a neural Turing machine, reasoning consistency is guaranteed through a generative adversarial network, model parameters are optimized in combination with a digital twin system and reinforcement learning, and a mental health service response is finally generated through a conditional generation model and hierarchical reinforcement learning. According to the method, the problem of asynchronism of multi-modal emotion feature dynamic evolution and context increment updating is effectively solved, accumulated drift of emotion state tracking is eliminated, and the continuity of reasoning logic is guaranteed.

The invention provides a point cloud completion method and system based on a high-dimensional feature field and structured tensor decomposition, and belongs to the field of artificial intelligence. Comprising the following steps: encoding geometric information carried by sparse point clouds through a space geometric encoder, and distributing the geometric information to three mutually orthogonal two-dimensional feature planes and a three-dimensional feature grid to jointly form a high-dimensional feature field in which space local information is reserved; for any query point in the space, combining the space coordinates of the query point and the local features sampled in the high-dimensional feature field as input, driving a decoder based on structured tensor decomposition, and reconstructing the geometric attribute value of the query point; a continuous geometric field function implicitly defines a three-dimensional surface after completion, and point cloud completion is completed by sampling the function and extracting a zero contour surface of the function. The technical problems that when sparse and incomplete point cloud data are processed, due to information bottleneck and lack of geometric priori, the complementation quality is poor, and macrostructures and microscopic details are difficult to consider at the same time are solved.

The invention discloses a dense monitoring Internet of Things remote estimation method based on multi-cell semantic enhancement, and belongs to the field of wireless communication. The method comprises the following steps: firstly, an edge device observes a reasoning target, encodes an observation value according to a semantic codebook, and transmits the observation value to an edge server; secondly, the edge server carries out vector quantization on the received signal and sends a quantized code word index to a cloud server through a forward link; and finally, the cloud server performs dequantization and joint decoding on the received quantization index to obtain an estimated value of the reasoning target in the multiple cells, and a remote estimation task is completed. In addition, a loss function is constructed based on an information bottleneck theory and a straight-through estimation gradient approximation method, and the system is trained to be optimal by adopting a two-stage training strategy. According to the method, semantic information extraction of the remote estimation task in a multi-cell scene is realized, the performance of the remote estimation task can be improved while code word redundancy is inhibited, and the utilization efficiency of spectrum resources is remarkably improved.

The invention relates to the technical field of social network information propagation prediction, in particular to an event propagation prediction method based on group influence. The method comprises the following steps: constructing a social network graph of target event participating users, calculating an activation probability matrix among the users by using historical propagation data, and performing social graph structure enhancement based on a threshold value to obtain a de-noised social graph; applying a graph neural network on the de-noised social graph and introducing information bottleneck constrained self-supervised contrast learning optimization node embedding, and reserving features useful for group division and propagation prediction; dynamically dividing user groups related to propagation according to the similarity and propagation context of de-noising node embedding to form a group-level propagation sequence; based on the group representation, utilizing an attention mechanism and a group relation graph to carry out modeling on propagation influence among the groups to obtain global group representation of the event; global group representation is combined with candidate user embedding, propagation probability distribution is calculated, and future participating users of corresponding events are predicted.

The invention discloses a multi-mode joint information source channel coding system for satellite-ground semantic communication. The multi-mode joint information source channel coding system comprises a joint information source channel coder and a decoder, the encoder is used for sequentially performing feature extraction, feature fusion, data cutting and phase modulation operation on input multi-modal data to generate a transmitting signal, and transmitting the transmitting signal to a ground base station through a channel; and the decoder is used for sequentially performing real and virtual part separation, data expansion, feature extraction and modal data recovery operation on the received signal so as to reconstruct multi-modal data. According to the system, public semantic information among multi-modal data can be mined, the time-varying low-signal-to-noise-ratio channel characteristics of a satellite downlink are combined, the information bottleneck theory highly related to data reconstruction quality is used as guidance to optimize a coding space, complex field transmission signals with the high compression ratio are generated in a self-adaptive mode, and the multi-modal multi-domain transmission signals are obtained. The reconstruction quality of the multi-modal data can be ensured, and the bandwidth and the energy consumption overhead of satellite communication can be remarkably reduced.

The invention discloses a concrete crack intelligent detection method based on four-cardinal-number coordinate attention enhancement, and belongs to the technical field of computer vision and civil engineering detection. According to the invention, a QuadCoord-YOLO network is constructed by improving a YOLO11 network architecture, and a CoT-CSP module and a QCE module are innovatively introduced: the CoT-CSP module deeply fuses a context converter mechanism and a CSP structure, so that the long-distance dependency relationship modeling capability is enhanced; the QCE module proposes a four-cardinal parallel channel attention mechanism for the first time, realizes three-path parallel multi-dimensional feature enhancement in combination with coordinate attention and deep convolution, and breaks through the information bottleneck of single-path compression of a traditional SE module. Experimental results show that according to the method, the bounding box detection mAP50 is improved from 72.6% to 78.1%, the segmentation mAP50 is improved from 58.9% to 66.6%, and only 3.2% of parameter quantity (82, 456 parameters) is increased. The method can be applied to an intelligent structure health monitoring system, and provides accurate technical support for concrete structure condition evaluation and maintenance decision.

The invention discloses a model optimization method and device based on hatred content classification. The method comprises the steps that text features and image features of expression package samples are obtained; based on the text features and the image features, multi-modal features of the expression package samples are obtained; using the multi-modal features of the plurality of expression package samples as training samples to train a deep learning model; optimizing a multi-modal distribution space of the deep learning model to enable the first probability distribution to be consistent with the second probability distribution; and on the basis of the optimized deep learning model, a hatred content classification model is obtained in combination with a variation information bottleneck mechanism. According to the method, emotional uncertainty caused by indirect expression is quantified by using a variance vector of a multi-modal feature in a deep learning model, and a multi-modal distribution space of the deep learning model is optimized, so that first probability distribution and second probability distribution are kept consistent; therefore, the recognition capability of the hatred content classification model on indirect expression of the hatred content is effectively improved.

The invention relates to an Internet of Vehicles cooperative sensing method based on semantic communication, and belongs to the technical field of intelligent connected automobiles. The method fuses a cooperative vehicle terminal, an own vehicle terminal and a wireless communication channel, and comprises the following steps: introducing an information bottleneck theory to the cooperative vehicle terminal to establish a variational optimization target, and extracting compact multi-view image features; mapping the multi-view image features into aerial view semantic features by using a semantic encoder based on a cross-view attention mechanism; mapping the extracted semantic features into a symbol stream through joint source channel coding, dynamically adjusting a coding strategy according to a channel state, and transmitting the coding strategy to an own vehicle through a wireless channel; semantic features are recovered by using a joint source channel decoder at an own vehicle end, positioning errors among multiple vehicles are corrected through a feature space calibration module, cross-vehicle feature alignment and fusion are realized by using an attention-based feature fusion module and a feature interaction precision enhancement module, and a global collaborative semantic segmentation map is generated. According to the method, compact features can be extracted through information bottleneck differential training so as to reduce transmission redundancy, anti-interference transmission under a time-varying channel is ensured by utilizing joint information source channel coding, the problem of spatial dislocation is solved through a multi-stage feature fusion mechanism, and the cooperative sensing precision under a low signal-to-noise ratio is remarkably improved.

The invention discloses a system for improving item recommendation accuracy based on space-time diffusion and information bottleneck. The system comprises the following modules: a normalization processing module, a space aggregation module, a time aggregation module, a time diffusion module, a space diffusion module, an information bottleneck module and a prediction module. The system can effectively model an error behavior which is not solved by a current next basket recommendation technology, and through an innovative spatial diffusion module, the system does not depend on strong association among commodities or historical preferences of users any more, but focuses on learning of article spatial combinatorial logic in a shopping basket, so that the system is more intelligent. Therefore, the system can identify the targeted and hidden combination strategy of the user, thereby improving the understanding of the real purchase intention of the user, and achieving the purpose of improving the article recommendation accuracy.

The invention relates to the field of network security, and discloses a network security situation awareness method and system based on deep learning, and the method comprises the following steps: obtaining multi-source network data; extracting time feature representation of nodes in the dynamic space-time diagram sequence through a multi-scale time feature extraction network; extracting spatial feature representation of the nodes through a graph neural network; fusing the time feature representation and the spatial feature representation; constructing a decoupling encoder based on the information bottleneck; and performing anomaly detection and attack path tracing based on the time hidden variable, the space hidden variable and the coupling hidden variable. By constructing a multi-scale time feature extraction network and introducing a dynamic receptive field adjustment mechanism, space-time coupling features are separated into a time hidden variable, a space hidden variable and a coupling hidden variable, and coupling information is comprehensively utilized during anomaly detection to improve detection precision. And an attack propagation path is accurately reconstructed based on pure spatial features during attack path tracing.

The application discloses a kind of high-impact weather condition-oriented wind power prediction information filtering method.It realizes the dynamic soft selection and reweighting of different input variables under different weather conditions through instance-level attention recalibration, improves the effective signal-to-noise ratio, and reduces the interference of noise channels on the learning process.At the same time, it removes irrelevant random disturbances from the information flow by variational information bottleneck compression regularization term, alleviates overfitting and noise memory problems caused by relying only on error minimization, and improves cross-condition generalization ability.Compared with the prior art, the present application can effectively filter out random disturbances irrelevant to the prediction task under the condition of strong noise and distribution fluctuation caused by high-impact weather, improve the stability, robustness and cross-scene generalization ability of the wind power prediction model, and have good engineering application value.

The application provides a trust feature completion and trust relationship prediction method and system for missing trust information, comprising: constructing a social relationship graph and a trust relationship graph to obtain social attribute features and trust attribute features respectively; compressing the social attribute features and the trust attribute features, and constructing a feature converter to convert the compressed social attribute features; constructing a user trust relationship prediction model, taking the compressed features of the user as input, and outputting a prediction label; taking the converted compressed social attribute features as input, and outputting a prediction label; taking a real label as supervision information, combining a reconstruction loss and an information bottleneck loss, training the model, and outputting a trust relationship. The application constructs a social behavior graph and a trust relationship graph, and learns the social attribute features and the trust attribute features of the user respectively. This multi-view learning method can more comprehensively capture the feature information of the user, and improve the accuracy of trust relationship prediction.

Systems and methods for particularly t-cell receptor complex optimization with reinforcement learning. Classifiers using variational information bottleneck with attention of experts (AVIB classifiers) can be fine-tuned (110) for different representations of desired t-cell receptor (TCR) sequences for a patient. Proximal policy optimization (PPO) models can be trained (120) with reinforcement learning using the AVIB classifiers as reward functions to achieve higher affinity in generating interaction sequences for the desired TCR sequences through automated decision making. The interaction sequences can be clustered (130) based on k-mer profiles to select the interaction sequences having highest binding scores in each cluster as final sequences. A biological functional potency of the final sequences can be validated (140).

The application discloses a kind of based on information bottleneck's debiasing recommendation method, comprising:1. Construct original data: user-product interaction matrix of user interaction to product, user biased attribute matrix;2. Using biased attribute encoder learns user biased representation in user-product interaction data;3. Based on deep graph neural network learns user and product representation matrix;4. Based on information bottleneck theory minimizes the mutual information between user representation and biased representation, user subgraph representation and biased representation, calculates loss function;5. Recomputes interaction matrix based on user and product representation matrix, calculates loss function;6. The loss function of joint step 4-step 5 is carried out information bottleneck learning, and model parameter is updated to model convergence.The application is based on the thought of information bottleneck, learns unbiased user representation under the demand of meeting recommendation task, so as to effectively reduce recommendation bias, guarantee the accuracy of recommendation.

The application discloses a supplier portrait visualization display method and system based on a knowledge graph, relates to the technical field of data processing, and comprises the following steps: performing information entropy driven compression on a pre-constructed supplier knowledge graph through a graph information bottleneck algorithm to generate a supplier compressed subgraph; obtaining a supplier feature set by positioning a target node in the supplier compressed subgraph; screening a target portrait template from a portrait template library according to the supplier feature set; inputting the supplier compressed subgraph, the supplier feature set and the target portrait template into a double-channel collaborative embedding architecture to generate a supplier portrait embedding vector; and associating the supplier portrait embedding vector with the target node for visualization display. The application is used to solve the problems of insufficient expression of supplier features, lack of adaptability of portrait templates and insufficient interpretability of visualization results in the prior art.

The invention discloses a vascular cognitive impairment prediction method, a computer program product and a terminal, and belongs to the field of deep learning. Feature information of medical time sequence data of a patient is extracted based on a time sequence data processing model; compressed feature representation of the feature information is learned through the information bottleneck; and performing prediction according to the compressed feature representation to obtain a prediction result of the disease progress. According to the method, the time sequence data processing model is used for capturing time dependence in the time sequence data, features related to the disease progression of dementia are effectively extracted from clinical data, then the features are screened by adopting information bottleneck, the most important features for classification prediction are identified, redundant and irrelevant information is eliminated, and the accuracy of classification prediction is improved. Therefore, the accuracy of classification prediction is improved. Furthermore, in order to simulate and generate various possible future data paths and reflect the uncertainty in the data, future process data is generated by using a diffusion model after feature extraction.

The invention belongs to the technical field of recommendation systems, and particularly relates to a recommendation method based on two-channel fusion of graph convolution and hypergraph convolution. According to the method, a two-channel architecture of graph convolution and hypergraph convolution is constructed and is respectively used for capturing explicit interaction information between users and merchants and high-order social information between the users, so that noise interference caused by hyperedge redundancy is effectively relieved; a gating attention fusion mechanism is designed, dynamic aggregation of double-channel embedding is achieved by means of a learnable weight distribution strategy, and recommendation accuracy and personalized expression ability are improved; merchant static features are introduced, category labels, geographical location information and the like are used as semantic vectors to improve recommendation performance, hyperedge noise propagation can be inhibited by means of an information bottleneck principle, and recommendation accuracy is improved.

The invention discloses a watch movement period health management system fusing multi-source sensing data, particularly relates to the technical field of data processing and health state management, and is used for solving the problem of insufficient iteration precision of a health assessment model caused by lack of key data dynamic screening and value weight adaptation capability in the prior art. Core operation parameters and associated environment data are acquired through a data acquisition module; the trigger judgment module judges whether a dynamic screening condition is triggered or not according to the data fluctuation index; when a condition is triggered, the key extraction module converts a time sequence into a symbol sequence and positions key data based on an information bottleneck theory; the causal purification module performs causal inference on the key data to peel off non-causal environmental interference; a weight adaptation module dynamically adapts a value weight to the purified data according to a movement wear mode; the model optimization module adopts weighted data to iteratively optimize a movement health assessment model; and intelligent purification of data value and adaptive enhancement of model precision are realized.

The invention provides a small molecule virtual screening method and device based on information bottleneck training, and relates to the technical field of artificial intelligence. The information bottleneck training-based small molecule virtual screening method comprises the following steps of: obtaining fusion vector representation of a drug small molecule sample and an undisturbed transcriptome data sample according to an identifier of the drug small molecule sample and the undisturbed transcriptome data sample by utilizing a drug small molecule pre-training model and an initial undisturbed encoder; forming a positive sample pair by the fusion vector representation and the prediction vector representation of the transcriptome variable quantity sample corresponding to the fusion vector representation, and forming a negative sample pair by the fusion vector representation and the prediction vector representation of the transcriptome variable quantity sample not corresponding to the fusion vector representation; and performing contrast loss training on the drug small molecule pre-training model, the initial transcriptome change encoder and the initial undisturbed encoder. According to the method, virtual screening of the drug small molecules can be accurately and robustly carried out.

The application discloses a kind of fusion knowledge enhancement and the method and system of original meaning prediction of confrontation training, it is related to natural language processing technical field, obtain target word and its corresponding original dictionary explanation;Original dictionary explanation is carried out semantic integrity evaluation using semantic perception selective enhancement intelligent agent;Target word, original dictionary explanation and supplementary context are structured as structured input sequence;Structured input sequence is sent into original meaning encoder, semantic features are extracted, and original meaning prediction vector is obtained;Using the method based on fast gradient, dynamic confrontation disturbance is applied to word embedding layer during model training process, clean sample loss and confrontation sample loss are optimized jointly to update model parameters;The original meaning label set corresponding to target word is output.The application introduces two big mechanisms of selective knowledge enhancement and confrontation training, and systematically solves the problem that decision boundary is weak due to static dictionary information bottleneck and original meaning long tail distribution in traditional original meaning prediction.

The application discloses a multimodal sentiment analysis method based on sentiment consistency distillation and information bottleneck, which comprises the following steps: respectively extracting sentiment features of image, text and audio modalities, and mapping the multimodal sentiment features into multimodal semantic vectors; then, decoupling the multimodal semantic vectors into modality-specific semantic vectors and modality-common semantic vectors; extracting sentiment consistency signals between modalities through an attention mechanism, and guiding the sentiment correlation information to be fused into the modality-common semantic vectors through a knowledge distillation mechanism; removing noise information in the modality-specific semantic vectors based on the information bottleneck theory; finally, splicing the optimized modality-specific semantic vectors and the modality-common semantic vectors into unified sentiment representation, and outputting corresponding sentiment categories through a classifier. The application can effectively alleviate the alignment deviation and information pollution problems caused by noise interference in multimodal data, improve the sentiment recognition accuracy, and take into account the calculation efficiency and model stability.

The invention provides a mobile crowdsourcing trust evaluation method and system, and relates to the technical field of mobile crowdsourcing. Comprising the following steps: constructing a hypergraph according to user nodes, working nodes and crowdsourcing tasks; constructing an initial trust evaluation model by using a heterogeneous graph neural network; constructing an optimization loss function by using a cross entropy loss function and a dual-information bottleneck loss function; performing parameter optimization on the initial trust evaluation model by utilizing an optimization loss function and a gradient descent method to obtain a final trust evaluation model; and performing trust prediction on the hypergraph by using the final trust evaluation model to obtain trust prediction values of the user and the working node corresponding to the specific crowdsourcing task. According to the method and the device, the problems that an existing trust relationship prediction scheme based on a graph neural network cannot fully consider environment information and a condition trust relationship under a specific task, so that the scheme cannot adapt to dynamics and diversity in a mobile crowdsourcing task, and the trust prediction accuracy is influenced are solved.

An image detection method based on embedded manifold representation and information bottleneck constraints is proposed. In the offline stage, a discriminator network and a generative flow model network (acting as an encoder or decoder) are initialized, and reconstruction loss function, information bottleneck loss function, and Jacobian matrix isometry loss function are calculated sequentially to obtain the total loss function. This allows for backpropagation training of both the generative flow model network and the discriminator network. In the online stage, the trained generative flow model network obtains discriminative and common features of the image to be tested. Image reconstruction and submanifold image reconstruction are performed only on images whose common features belong to the common feature space. Image classification is achieved by calculating the distance between the reconstructed image and the reconstructed images of each category's submanifold. This invention avoids meaningless calculations on invalid samples by pre-analyzing the images to be classified, saving computational resources. Furthermore, by utilizing the guidance of embedded manifold representation algorithm in image sample projection and information bottleneck constraints in feature extraction, it demonstrates excellent classification robustness.

Method (200) for recognizing visually matching tones, wherein the method comprises: Receiving visual training data (105) at a visual coder (110) that has an initial machine learning (ML) model; Identifying data corresponding to a visual object in the visual training data (105) using the first ML model; Receiving audio training data (107) synchronized with the visual training data at an audio forwarding regulator (115) which has a second ML model, wherein the audio training data (107) has a visually matching tone and a visually mismatched tone, both of which are synchronized with one and the same frame in the visual training data (105) which contains the visual object, wherein the visually matching tone corresponds to the visual object, whereas the visually mismatched tone is generated by a sound source which is not visible in the same frame; Filtering data matching the visually appropriate tone from an output of the second ML model using an information bottleneck (120); and Training a third ML model following the first and second ML models (235) using the data corresponding to the visual object and data corresponding to the visually inappropriate tone.