Jiangnan University

Robert Boschgmbh

Wei Fang

Jun Sun

Xiaojun Wu

Hengyang Lu

Xin Zhang

Shitong Wang

Lin Wang

Chao Fan

Chenyou Fan

Pengjiang Qian

Runsu Zhu

Jian Yao

Peiming Ren

Yi Gu

Zhihuang Wang

Yizhang Jiang

Aiguo Chen

He Xu

Xingguo Zhang

The invention provides a method for constructing a target prediction model in a multicenter small sample scenario and a prediction method. In combination with the idea of transfer learning, a training set of a new node is predicted by directly using knowledge of a trained node, and a prediction error sample is used to reflect a difference between the new node and the trained node. The difference is used as supplementary knowledge. Model knowledge of the new node is quickly acquired, to avoid training of a new node from scratch each time. Finally, parallel integration of incremental subclassifiers is implemented by using a ridge regression method, so that the deployment time and costs are greatly reduced. The generalization of models is ensured through sharing of historical knowledge and a knowledge discarding mechanism, good classification effect can also be achieved for a node with a small sample size.

Method for constructing target prediction model in multicenter small sample scenario and prediction method

The present invention provides a text classification backdoor attack method, system, device and a computer storage medium. The method includes: training a pretraining model by using a clean training set to obtain a clean model; generating a pseudo label data set by using a positioning label generator; performing multi-task training on a Sequence-to-Sequence model by using the pseudo label data set to obtain a locator model; generating a backdoor data set by using the locator model; and training the clean model by using the backdoor data set to obtain a dirty model. A pseudo label data set is generated by using a pretrained clean model without manual annotation. A backdoor attack location in a text sequence may be dynamically predicted by using a locator model based on a Sequence-to-Sequence and multi-task learning architecture without manual intervention, and a performance indicator obtained by dynamically selecting an attack location is better.

Text classification backdoor attack prediction method, system, and device

The present disclosure discloses a deep spatial-temporal similarity method for air quality prediction, and belongs to the technical field of environmental protection. When the method predicts air quality-related indexes of a target site, a temporal change of air pollution and a spatial diffusion relationship are effectively combined, and then spatial-temporal similarity sites of the target site are selected; air quality monitoring data collected by the target site, the spatial-temporal similarity site of the target site and geographical neighbour sites of the target site and meteorological data are respectively taken as inputs of a long short term memory network (LSTM) model to obtain uncorrelated output results, and then predicted values of air quality-related index data of the target site are obtained in a mode of support vector regression (SVR) integration. The present disclosure effectively combines the temporal change of air pollution with the spatial diffusion relationship, and namely proposes a more efficient way to select more highly relevant data to predict air quality so that a prediction result is more accurate.

Deep spatial-temporal similarity method for air quality prediction

Disclosed is a real-time object detection method deployed on a platform with limited computing resources, which belongs to the field of deep learning and image processing. In the present invention, YOLO-v3-tiny neural network is improved, Tinier-YOLO reserves the front five convolutional layers and pooling layers of YOLO-v3-tiny and makes prediction at two different scales. Fire modules in SqueezeNet, 1&#xd7;1 bottleneck layers, and dense connection are introduced, so that the structure is used to achieve smaller, faster, and more lightweight network that can be run in real time on an embedded AI platform. The model size of Tinier-YOLO in the present invention is only 7.9 MB, which is only &#xbc; of 34.9 MB of YOLO-v3-tiny, and &#x215b; of YOLO-v2-tiny. The reduction in the model size of Tinier-YOLO does not affect real-time performance and accuracy of Tinier-YOLO. Real-time performance of Tinier-YOLO in the present invention is 21.8% higher than that of YOLO-v3-tiny and 70.8% higher than that of YOLO-v2-tiny. Compared with YOLO-v3-tiny, accuracy of Tinier-YOLO is increased by 10.1%. Compared with YOLO-v2-tiny, accuracy of Tinier-YOLO is increased by nearly 18.2%. Tinier-YOLO in the present invention can still achieve real-time detection on a platform with limited resources, and effects are better.

Real-time target detection method deployed on platform with limited computing resources

Growing community of inventors

Wuxi, China

Wei Fang