36 results about "Information bottleneck method" patented technology

Some embodiments of the invention provide a novel method for training a quantized machine-trained network. Some embodiments provide a method of scaling a feature map of a pre-trained floating-point neural network in order to match the range of output values provided by quantized activations in a quantized neural network. A quantization function is modified, in some embodiments, to be differentiable to fix the mismatch between the loss function computed in forward propagation and the loss gradient used in backward propagation. Variational information bottleneck, in some embodiments, is incorporated to train the network to be insensitive to multiplicative noise applied to each channel. In some embodiments, channels that finish training with large noise, for example, exceeding 100%, are pruned.

Some embodiments of the invention provide a novel method for training a quantized machine-trained network. Some embodiments provide a method of scaling a feature map of a pre-trained floating-point neural network in order to match the range of output values provided by quantized activations in a quantized neural network. A quantization function is modified, in some embodiments, to be differentiable to fix the mismatch between the loss function computed in forward propagation and the loss gradient used in backward propagation. Variational information bottleneck, in some embodiments, is incorporated to train the network to be insensitive to multiplicative noise applied to each channel. In some embodiments, channels that finish training with large noise, for example, exceeding 100%, are pruned.

The invention discloses a probability domain generalization learning method based on meta-learning, belongs to the field of meta-learning, and aims to combine a meta-learning thought into domain generalization for the first time and solve the problem that parameters are linearly increased along with increase of the number of source domains in domain generalization by utilizing a meta-learning framework. The variational information bottleneck idea is combined into meta-learning and domain generalization for the first time, so that the generalization ability of the patent can be further improved; according to the method, the problem that parameters linearly increase along with the number of source domains can be solved through meta-learning; according to the technical scheme, the probabilitydomain generalization learning method based on meta-learning is formed by combining the variational thought with the information bottleneck and fusing the variational thought and the information bottleneck into a unified probability framework, and the brand-new and effective probability domain generalization learning method based on meta-learning can be formed by combining the variational thoughtwith the information bottleneck.

The invention discloses an unsupervised domain adaptive method for beneficial feature alignment under class conditions, and the method comprises the steps of calculating all source domain images and target domain images, and obtaining a pseudo tag of the target domain image; decoupling the pseudo labels of the source domain image and the target domain image through variational information bottlenecks, filtering features irrelevant to tasks out, and obtaining beneficial and migratable features; estimating intra-class difference and inter-class spacing by using a conditional slice Wharisstein distance, minimizing the intra-class difference and maximizing the inter-class spacing in a cross-domain manner, reducing class-level distribution difference between a source domain and a target domain, and obtaining domain-invariant discriminant features. According to the method, decoupling of a source domain and a target domain can be achieved, class-level information is embedded into the slice Weiisstein distance, beneficial feature alignment is achieved, and meanwhile beneficial feature decoupling and class condition feature alignment are achieved so as to promote better domain adaptation.

The invention belongs to the field of natural language processing and sentiment analysis, and particularly relates to a real estate market analysis method and device based on deep transfer learning, and equipment, and the real estate market analysis method comprises the steps: collecting real estate network public opinion data, and carrying out the preprocessing of the public opinion data; constructing a deep multi-channel neural network integrated with a variational information bottleneck; pre-training the network by using a large amount of annotation data in the related field; finely adjusting a pre-established network by using a small amount of marked public opinion data in a transfer learning mode; and performing emotional tendency analysis on unlabeled real estate public opinion dataon the migrated network, and obtaining a final real estate market emotional analysis result. According to the real estate market analysis method, deep migration learning and real estate network publicopinions are combined, and real estate market emotion can be accurately analyzed, so that reference and guidance are provided for policy making of related departments, decision deployment of companies and enterprises and investment planning of individual consumers.

The invention provides a three-dimensional fuzzy clustering method based on an information bottleneck theory. The three-dimensional fuzzy clustering method provided by the invention can be used for clustering the user-searching-document three-dimensional data. After the three-dimensional fuzzy clustering method is used, the simultaneous clustering of three dimensions can be realized for a three-dimensional contingency table, and the clustering accurate rate can be effectively enhanced on the basis of the distance between an information bottleneck theory measurement document and a cluster centroid.

The invention belongs to the technical field of medical image processing, and provides a breast cancer image information bottleneck multi-task classification and segmentation method and system. The method comprises the following steps: acquiring a plurality of breast images of contrast enhanced X-ray photography and corresponding benign and malignant categories and lump position pixel-level labels; performing preprocessing operation on each acquired breast image of the contrast enhanced X-ray photography; adopting a multi-task network to extract multi-task shared representation for each preprocessed mammary gland image; inputting the shared representation to a classification encoder to obtain an encoding tensor of an intermediate layer, and then sending the intermediate encoding tensor to an information bottleneck attribution module for feature compression and benign and malignant classification to obtain a classification task tensor; adapting the shared representation to a segmentation network to obtain a segmentation task tensor; and carrying out feature fusion on the classification task tensor and the segmentation task tensor to segment the focus.

The invention discloses a safety assessment framework building method for deep learning bypass analysis. The method comprises the steps that a secret key information quantum framework based on modeling output and a deep learning sub-framework composed of cross entropy and a loss function are included; and the bypass analysis sub-framework consists of a condition guessing entropy, a derivative guessing entropy and a success rate. According to information bottleneck theoretical analysis, perception information from the last hidden layer to the output layer of the DNN is calculated through probability distribution to serve as a quantitative evaluation index of the DLSCA, and experiments prove that the index is more accurate in measurement compared with an existing quantitative evaluation index and has lower requirements for the order of magnitude of data.

The invention relates to the field of artificial intelligence, and particularly discloses a voice classification model training method and device, equipment and a storage medium, and the method comprises the steps: obtaining sample data which comprises a sample voice and a sample label corresponding to the sample voice; preprocessing the sample voice to obtain a vector matrix corresponding to the sample voice; inputting the vector matrix into a variational information bottleneck processing network of an initial voice classification model to obtain sentence representation corresponding to the sample voice; and inputting the sentence representation into a classification network of the initial voice classification model to obtain a prediction label, and then performing iterative training on the initial voice classification model according to the sample label and the prediction label to obtain a trained voice classification model. A variational information bottleneck processing network is added into a voice classification model, so that sample information is compressed, useless information is inhibited, and an over-fitting phenomenon is reduced.

The invention provides a microblog hot topic detection method and apparatus and relates to the technical field of network information digging. Difficulty for detection of hot topics of highly-sparse microblog texts can be eliminated. The method comprises steps of extracting meaningful strings in an information flow and conducting heat analysis to the meaning string to screen hotspot features, combining hotspot features according to information bottleneck theory, calculating combining degree indication parameters after the combination, and determining the hot topic based on the combining degree indication parameters.

The invention designs an online learning input degree identification method and system based on deep learning, and the method comprises the steps: firstly, carrying out the student face detection through YOLOv4 in order to guarantee that an image is not affected by an irrelevant background; secondly, aiming at the problems that the VGG16 network is huge in parameter quantity, time-consuming in training and the like, an improved VGG16 model is provided, and meanwhile, in the model training process, a depth deterministic information bottleneck method DIB is adopted to make up for the deficiency of a traditional loss function so as to obtain relatively compact feature expression, reduce generalization errors and improve the universality and stability of the model, and achieving precise identification of the learning input degree in a complex online learning scene; and finally, through comparison and analysis with various methods such as traditional machine learning and other deep learning, the effectiveness of the method is verified.

The invention relates to the technical field of artificial intelligence, and provides a named entity recognition method, device and equipment and a computer readable storage medium, and the named entity recognition method comprises the steps: obtaining a pre-trained named entity recognition model, obtaining a to-be-recognized first statement, and inputting the to-be-recognized first statement into the named entity recognition model, performing the following named entity recognition processing by using the mission name entity recognition model: performing word segmentation processing on the first statement to obtain a second statement comprising a plurality of segmented words; performing feature extraction on the plurality of split words to obtain a plurality of word embedding feature vectors; processing the second statement according to the plurality of word embedding feature vectors to obtain a plurality of cross-domain information features; processing the plurality of cross-domain information features through an information bottleneck layer to obtain a plurality of information bottleneck features; the classification function is adopted to classify and recognize the multiple information bottleneck features, the corresponding named entity category is determined, the unregistered words in the named entity can be better recognized, and the named entity recognition accuracy is improved.