Hong Kong Applied Science and Technology Research Institute Company, Limited

Zhi Bin Lei

Yang Liu

Chao Feng

Kangheng Wu

Man Yau Chiu

Hailiang Li

Yi Ping Tse

Yu Keung Ng

Yan Wang

Zhixi Li

Lawrence Kwan Yeung

Zuyao Wang

Wei Chen Huang

Zhengmairuo Gan

Cheuk Hang Chiu

Ji Hyoun Park

Yi Xiang

Vincent Wai-Leuk Tam

A method of high dimensional data analysis in real-time comprising executing dimension-reducing an input historical data set under a t-SNE model and determining from the resulting dimension-reduced data set a recent; further dimension-reducing the recent group data set under a PCA model; statistical analyzing the further dimension-reduced data set to determine a threshold group for distinguishing abnormal data from normal ones in a real-time data stream. The method may further include training a classifier using the abnormal or normal data set for predicting anomaly in the real-time data source system. Alternatively, a discrepancy training data set is computed from one of the normal and abnormal data sets and be used to train one of independent normal and abnormal data regression models; with the other one trained by transfer learning based on the trained one. The trained regression models are then used to predict discrepancy values.

Apparatus and method of high dimensional data analysis in real-time

A method for summarizing text with sentence extraction including steps as follows. Sentences are extracted from a document including text by a natural language processing (NLP) based feature extractor. A word vector set with respect to each of the sentences is generated by a processor. The word vector set with respect to each of the sentences is used to generate a n-grams vector set and a phrase-n vector set with respect to each of the sentences. A word score representing similarity between the word vector sets, a n-grams score representing similarity between the n-grams vector sets, and a phrase-n score representing similarity between the phrase-n vector sets are computed. The word, n-grams, and phrase-n scores are combined to compute an edge score. Text features are selected from the sentences using the edge scores of the sentences, so as to output a summary of the document.

Method of summarizing text with sentence extraction

A method for applying image encoding recognition in the execution of natural language processing (NLP) tasks, comprising the processing steps as follows. A sentence from a textual source is extracted by an NLP-based feature extractor. A word vector is generated in response to the sentence by the NLP-based feature extractor. The word vector is converted into a feature vector {right arrow over (b)} by the NLP-based feature extractor, in which the feature vector {right arrow over (b)} satisfies {right arrow over (b)}&#x2208;and the parameter m is a positive integer. The feature vector is transformed into an image set having a plurality of two-dimensional images by a transformer. The image set is fed to a neural network to execute image recognition by a processor, so as to analyze the sentence.

Apparatus and method for applying image encoding recognition in natural language processing

A method for annotating a batch of original samples is provided. A first subset of original samples, selected from the batch and determined by minimizing an entropy-mean difference between the first subset and the batch, is used for human annotation to yield human-annotated samples. The human-annotated samples are used as training data to configure an annotation process for annotating an input sample to yield an annotated output sample, and a check process for verifying annotation accuracy of the annotated output sample. Remaining original samples in the hatch are processed by the annotation process to yield machine-annotated samples, whose accuracy is verified by the check process. In one embodiment, part of the original samples corresponding to erroneous machine-annotated samples are selected for human annotation. Resultant additional human-annotated samples are used to update the two processes. The remaining original samples not yet annotated are then processed by the two processes.

Efficient annotation of large sample group

A method for recognizing and extracting named entities in a natural language input text processing comprising: performing, by a compressed named entities recognition (NER)-model-based named entity recognizer, a first stage NER on the input text to generate a first stage determination of whether at least one named entity exists in the input text; if the first stage NER determines no named entity exists in the input text, performing, by a rule based named entity recognizer, a second stage NER on the input text to generate a second stage NER result; if the first stage NER determines at least one named entity exists in the input text, generating, by the compressed NER-model-based named entity recognizer, a first stage NER result; and integrating, by a NER result integrator, the first stage NER result and the second stage NER result to generate a final NER result.

Efficient and accurate named entity recognition method and apparatus

Methods and systems which provide super-resolution synthesis based on weighted results from a random forest classifier are described. Embodiments apply a trained random forest classifier to low-resolution patches generated from the low-resolution input image to classify the low-resolution input patches. As each low-resolution patch is fed into the random forest classifier, each decision tree in the random forest classifier 'votes' for a particular class for each of the low-resolution patches. Each class is associated with a projection matrix. The projection matrices output by the decision trees are combined by a weighted average to calculate an overall projection matrix corresponding to the random forest classifier output, which is used to calculate a high-resolution patch for each low-resolution patch. The high-resolution patches are combined to generate a synthesized high-resolution image corresponding to the low-resolution input image.

Systems and methods for super-resolution synthesis based on weighted results from a random forest classifier

A method for organizing and processing feature based data structures that can be used in linguistic spell checking and auto-correction, comprising: splitting an original dictionary into sub-dictionaries based on different values of a common feature such as high frequency words; receiving an input text that contains errors; determining a sub-dictionary selection feature from the input human-readable text; selecting the sub-dictionary based on the determined sub-dictionary selection feature; executing a first matching in the selected sub-dictionary, wherein a match is found if a similarity between the characters, words, or phrases in proximity of the errors in the input text and a character, word, or phrase in the sub-dictionary is above a threshold; if a unique match is found, the result is returned as an output to correct the errors; otherwise, executing a second matching with a raised threshold, and repeating the second matching until a unique match is found.

System and method for organizing and processing feature based data structures

A method for a vehicle management system with a perspective view camera generating perspective images of vehicles in a traffic direction, including performing an anti-perspective transform on the perspective images; separating each anti-perspective image into sub-images along one of a radial direction and a tangential direction relative to the traffic direction; determining a scale factor for each sub-image based upon measuring a scale of each vehicle at plural positions in one image using a frame difference method performing a scale transform for each sub-image using the corresponding scale factors for the sub-image; combining each of the scale transformed sub-images for each vehicle into corresponding enhanced anti-perspective images; performing vehicle detection for each vehicle based a combination of the enhanced anti-perspective images in which tracking is enhanced with an optimized detection box size range determined by the enhanced anti-perspective images; and performing vehicle tracking for each vehicle based on the combination of the enhanced anti-perspective images in which detecting is enhanced with the optimized detection box size range determined by the enhanced anti-perspective images.

Vehicle detection, tracking and localization based on enhanced anti-perspective transformation

A computer implemented method of load shedding used in stream computing system that considers the relative importance of each of the applications processing the incoming input data or events. The method of load shedding method also accounts for system physical constraints, such as memories and CPU utilization. The load shedding method first observes the workload of each application and arriving rate of the incoming input data or events. If the system is under an overloading condition, calculate a input data or event drop ratio for each application such that the projected sum of all applications' workload will be at or below the system capacity when the unprocessed input data or events are dropped according to the drop ratio for each application.

Configurable dynamic load shedding method in distributed stream computing system

This invention discloses an encoding apparatus and a decoding apparatus for scalable video coding such that a non-scalable video decoder is usable for decoding a scalable video bit-stream comprising a base layer bit-stream and an enhancement layer bit-stream. In one embodiment, a source video frame is downscaled to give a downscaled video frame, which is then encoded into the base layer bit-stream. The difference between the source video frame and an up-scaled video frame reconstructed from the downscaled video frame in the base layer bit-stream yields a residual frame. The residual frame is partitioned into a number of residual sub-frames each having a resolution that is the downscaled video frame's resolution. The residual sub-frames are encoded into the enhancement layer bit-stream. Thereby, a non-scalable encoder is usable to encode both the downscaled video frame and the residual sub-frames, allowing both bit-streams to be decodable by employing only one non-scalable decoder.

Growing community of inventors

Hong Kong, China

Zhi Bin Lei