Abstract: Embodiments of the present disclosure relate to a method, a device, and a computer program product for model arrangement. The method includes determining a target model for processing data. The method further includes dividing the target model into a plurality of modules that implement different tasks. The method further includes determining a quantity of parameters of a target module in the plurality of modules and a size of transmission data related to the target module. The method further includes determining an arrangement position of the target module based on the quantity and the size. With this method, the amount of data transmitted can be minimized, and the computing time and the presentation time of information presented to users can be reduced, thereby improving the user experience.

Abstract: A method in an illustrative embodiment includes determining, by a first deep learning model on a storage device, whether a fault is caused by an environmental factor. The method further comprises: determining, by a second deep learning model on the storage device in response to determining that the fault is caused by the environmental factor, whether the fault can be solved locally in the storage device. The method further comprises: processing the fault according to a knowledge base in response to determining that the fault can be solved locally in the storage device. The method further comprises: sending the fault to a third deep learning model on a cloud device in response to determining that the fault cannot be solved locally in the storage device, the first deep learning model and the second deep learning model being obtained by model distillation of the third deep learning model.

Abstract: Embodiments of the present disclosure provide a method, an electronic device, and a computer program product for information processing. In an information processing method, a first network state representation and a first content request event of an emulated network are provided from an emulator to an agent for reinforcement learning, wherein the first content request event indicates that a request node in the emulated network requests target content stored in a source node. The emulator receives first action information from the agent, wherein the first action information indicates a first caching action determined by the agent, the first caching action including caching the target content in at least one caching node between the request node and the source node. The emulator collects, based on the execution of the first caching action in the emulated network, first training data for training the agent.

Abstract: Embodiments of the present disclosure relate to a method, a device, and a computer program product for verifying a classification result. The method comprises: obtaining a plurality of clusters for training samples by clustering feature representations of the training samples using labels of the training samples; determining, based on the plurality of clusters, a clustering class to which an input image belongs; acquiring a classification result of a classification model for the input image; and verifying the degree of reliability of the classification result for the input image based on the clustering class and the classification result. In this manner, the verification according to embodiments of the present disclosure not only easily combines domain-specific knowledge and improves the detection precision, but also saves computational overhead and storage resources, thus enabling the solution to be deployed in edge devices or Internet of Things devices with limited computational power.

Abstract: Embodiments of the present disclosure relate to a method, an electronic device, and a computer program product for generating an avatar. The method includes generating an indication of correlation among image information, audio information, and text information of a video. The method may further include generating, based on the indication of the correlation, a first feature set and a second feature set representing features of a target object in the video, wherein the first feature set represents invariant features of the target object in the video, and the second feature set represents equivariant features of the target object in the video. The method may further include generating the avatar based on the first feature set and the second feature set. With this method, the generated avatar can be made more accurate and vivid with a better effect, while also reducing data annotation cost, improving operation efficiency, and enhancing user experience.

Abstract: Embodiments of the present disclosure provide a method, an electronic device, and a computer program product for speech synthesis. The method for speech synthesis includes: extracting a plurality of voice feature vectors of a plurality of speakers from a plurality of audios corresponding to the plurality of speakers; calculating a first loss function based on distances between the plurality of voice feature vectors of the plurality of speakers; calculating a second loss function according to a plurality of texts and a plurality of corresponding real audios; and generating a speech synthesis model based on the first loss function and the second loss function. By implementing the method, the speech synthesis model can be optimized and trained, so that a high-quality audio with target voice features can be outputted based on the texts.

Abstract: Embodiments of the present disclosure relate to a method, an electronic device, and a computer program product for acquiring an image. The method includes distilling an original image set through a capsule neural network model to generate a distilled image set, wherein the distilled image set includes a plurality of distilled images. The method further includes acquiring a first feature of a first image through the capsule neural network model. The method further includes acquiring a plurality of distilling features of the plurality of distilled images respectively through the capsule neural network model. The method further includes determining a plurality of similarities between the first feature and the plurality of distilling features respectively. The method further includes acquiring at least one original image matching the first image based on the plurality of similarities.

Abstract: Embodiments disclosed herein include a method, an electronic device, and a computer program product for target image processing. The method includes receiving a target image and generating a first Shapley value for a feature of the target image based on the received target image. The method further includes sending, in response to satisfying a predetermined condition, a request for acquiring a second Shapley value to a cloud server. The method further includes receiving the second Shapley value for a latent feature of the target image from the cloud server, where the second Shapley value is more accurate than the first Shapley value. In some embodiments, through joint collaboration between a terminal device such as an edge device and a cloud server, rapid calculation of a Shapley value can be achieved at the terminal device, and accurate calculation of a Shapley value can be achieved at the cloud server.

Abstract: Embodiments of the present disclosure relate to a method, a device, and a computer program product for text to speech. The method includes encoding a reference waveform of a first speaker to obtain an encoded style feature separated from a second speaker. The method further includes transferring the encoded style feature to a spectrogram obtained by encoding an input text, to obtain a style transferred spectrogram. The method further includes converting the style transferred spectrogram into a time-domain speech waveform. According to the method for text to speech in the present disclosure, a comparative learning framework can also flexibly and effectively synthesize speech with a style of a target speaker, thus realizing lightweight speech style transfer, making it possible to learn high-quality and recognizable features of speech synthesis, and realizing effective speaker feature learning. In addition, the model will be beneficial to other downstream tasks.

Abstract: Embodiments of the present disclosure provide a method, an electronic device, and a computer program product for data processing. A method for data processing includes acquiring an image input and text input pair, the image input and text input pair comprising image input and text input, obtaining an image feature codebook corresponding to the image input and a text feature codebook corresponding to the text input, and training a text-image association model by using the image feature codebook and the text feature codebook, the text-image association model implementing the data processing based on the association between the image input and the text input. For example, a model may be adopted to learn an association between an image feature codebook corresponding to image input and a text feature codebook corresponding to text input, and data processing, such as data compression or data retrieval, is implemented using the learned association.

Abstract: Methods, devices and computer program products for processing video are disclosed herein. A method includes: generating, based on a reference image and a first frame of a video comprising an object, a two-dimensional avatar image of the object; and generating a base three-dimensional avatar of the object by performing a three-dimensional transformation on the two-dimensional avatar image and the object in the first frame. The method further includes: generating a three-dimensional avatar video corresponding to the video based on the base three-dimensional avatar and features of the video, the features comprising differences of the object between adjacent frames of the video. This solution enables the generation of a customized three-dimensional avatar video 10 for an object in a video, where the avatar can move in synchronization with the object and retain the unique features of the object, and can provide a more detailed and vivid representation than a two-dimensional avatar.

Abstract: A method in one embodiment includes: extracting first image features of a first image of a first resolution and second image features of a second image of a second resolution, wherein the first resolution is less than the second resolution, and the first image and the second image correspond to each other. The method further includes: extracting reference image features of a reference image, wherein the reference image includes edges and modes for reconstruction reference. The method further includes: generating a third image of a third resolution based on the first image features, the second image features, and the reference image features, wherein the third resolution is less than or equal to the second resolution. By using this method, super-resolution image processing can be implemented on a low-resolution image to obtain a reconstructed image with similar or the same resolution as the original image, with reduced storage space and device costs.

Abstract: Embodiments of the present disclosure relate to a method, an electronic device, and a computer program product for training a data classification model. The method includes generating a first training rule based on probabilities of classifying a plurality of sample data into corresponding classes by a data classification model. The method also includes generating a second training rule based on relevances of the plurality of sample data to the corresponding classes. In addition, the method also includes training the data classification model using the first training rule and the second training rule. With this method, a data classification model is trained, so that the data classification accuracy of the data classification model and the robustness to noise can be improved.

Abstract: A method in one embodiment includes receiving, at an edge device, new data for training a model, the edge device having stored distilled data used to represent historical data to train the model, the historical data being stored in a remote device, and the amount of the historical data being greater than the amount of the distilled data. The method further includes training the model based on the new data and the distilled data. With the data processing solution of this embodiment, the model can be trained at the edge device with fewer storage resources based on the distilled data, thereby achieving higher model accuracy.

Abstract: Embodiments of the present disclosure provide a method, an electronic device, and a computer program product for information processing. In an information processing method, based on multiple weights corresponding to multiple words in text, a computing device determines a target object associated with the text among predetermined multiple objects, and also determines, among the multiple words, a set of key words with respect to the determination of the target object. Next, the computing device determines, among the set of key words, a set of target words related to a text topic of the text. Then, the computing device outputs the set of target words and an identifier of the target object in an associated manner. In this way, the credibility of the target object associated with the text that is determined by the information processing method is improved, thereby improving the user experience of the information processing method.

Abstract: Embodiments of the present disclosure relate to a computer-implemented method, device, and computer program product. The method includes: determining, based on a set of user behavior data over a first time period, a set of behavioral features for a user behavior over the first time period. The method further includes: determining, based on the set of behavioral features, a set of anomaly scores for the user behavior according to an anomaly detection model. The anomaly detection model is trained based on user behavior data over a second time period. The method further includes: updating a previously determined anomaly score threshold based on comparison of the determined set of anomaly scores with the anomaly score threshold. The anomaly score threshold is used for indicating whether the user behavior is anomalous. By continuously updating the anomaly score threshold, the solution uses the updated anomaly score threshold for anomaly detection of user behavior.

Abstract: A method in an illustrative embodiment of the present disclosure includes determining, utilizing a first diagnosis model deployed in a storage system, whether a cause of a fault belongs to environmental factors. The method further includes determining, responsive to determining that the cause of the fault belongs to the environmental factors, whether the fault can be solved locally in the storage system. The method further includes sending, responsive to determining that the fault cannot be solved locally in the storage system, the fault to a second diagnosis model, wherein the first diagnosis model is obtained by distilling the second diagnosis model. According to the method for fault diagnosis of the present disclosure, particular faults can be diagnosed and solved locally in a storage system, so that the workload of a customer support team of the storage system in a cloud can be reduced.

Abstract: Embodiments of the present disclosure relate to a method, a device, and a computer program product for model arrangement. The method includes determining a target model for processing data. The method further includes dividing the target model into a plurality of modules that implement different tasks. The method further includes determining a quantity of parameters of a target module in the plurality of modules and a size of transmission data related to the target module. The method further includes determining an arrangement position of the target module based on the quantity and the size. With this method, the amount of data transmitted can be minimized, and the computing time and the presentation time of information presented to users can be reduced, thereby improving the user experience.

Abstract: A method includes generating a feature map of a second resolution for an input image of a first resolution, the first resolution being less than the second resolution. The method further includes generating a feature map of the second resolution and a feature map of the first resolution through a first residual-based dense back projection (RDBP) network. In addition, the method further includes generating a feature map of the second resolution and a feature map of the first resolution through a second RDBP network, the first RDBP network and the second RDBP network respectively including a downsampling back projection layer, an upsampling back projection layer, and an attention layer. The method further includes generating an output image of the second resolution based on the feature maps of the second resolution and the feature maps of the first resolution generated through the first RDBP network and the second RDBP network.

Abstract: Embodiments of the present disclosure provide a method, an electronic device, and a computer program product for detecting model performance. The method may include acquiring a prediction result of an input feature using a target model to determine a confidence of the prediction result. The method may further include reconstructing the input feature using a self-coding model to determine a reconstruction error, the reconstruction error being a difference between the input feature before being reconstructed by the self-coding model and the input feature after being reconstructed by the self-coding model. In addition, the method may include determining a detection result of the target model at least based on a comparison between the confidence and a first threshold and a comparison between the reconstruction error and a second threshold.