This study is based on the ICASSP 2025 Signal Processing Grand Challenge’s Accelerometer-Based Person-in-Bed Detection Challenge, which aims to determine bed occupancy using accelerometer signals. The task is divided into two tracks: "in bed" and "not in bed" segmented detection and streaming detection, facing challenges such as individual differences, posture variations, and external disturbances. We propose a spectral-temporal fusion-based feature representation method with mixup data augmentation, and adopt Intersection over Union (IoU) loss to optimize detection accuracy. In the two tracks, our method achieved outstanding results of 100.00% and 95.55% in detection scores, securing first place and third place, respectively.

This study focuses on the First VoicePrivacy Attacker Challenge within the ICASSP 2025 Signal Processing Grand Challenge, which aims to develop speaker verification systems capable of determining whether two anonymized speech signals are from the same speaker. However, differences between feature distributions of original and anonymized speech complicate this task. To address this challenge, we propose an attacker system that combines Data Augmentation enhanced feature representation and Speaker Identity Difference enhanced classifier to improve verification performance, termed DA-SID. Specifically, data augmentation strategies (i.e., data fusion and SpecAugment) are utilized to mitigate feature distribution gaps, while probabilistic linear discriminant analysis (PLDA) is employed to further enhance speaker identity difference. Our system significantly outperforms the baseline, demonstrating exceptional effectiveness and robustness against various voice anonymization systems, ultimately securing a top-5 ranking in the challenge.

Event Extraction (EE), aiming to identify and classify event triggers and arguments from event mentions, has benefited from pre-trained language models (PLMs). However, existing PLM-based methods ignore the information of trigger/argument fields, which is crucial for understanding event schemas. To this end, we propose a Probabilistic reCoupling model enhanced Event extraction framework (ProCE). Specifically, we first model the syntactic-related event fields as probabilistic biases, to clarify the event fields from ambiguous entanglement. Furthermore, considering multiple occurrences of the same triggers/arguments in EE, we explore probabilistic interaction strategies among multiple fields of the same triggers/arguments, to recouple the corresponding clarified distributions and capture more latent information fields. Experiments on EE datasets demonstrate the effectiveness and generalization of our proposed approach.

Landslide natural disasters (LND) have high frequency, wide distribution, and multiple occurrences, causing significant losses to personal and property safety. LNDs account for over 70% of natural geological disasters in China, often caused by precipitation. Xingguo County, Jiangxi Province, is prone to LND due to its geographical location. Rainfall-induced LNDs account for over 70% of the county's LND. In this study, a digital modeling and machine learning approach is used to evaluate the susceptibility of rain-induced landslides in Xingguo County and generate a high-precision susceptibility map. Six influence factors are selected, and four machine learning algorithms, including support vector machine (SVM), decision tree (DT), back propagation neural network (BPNN), and random forests (RF), are used for susceptibility evaluation. A rainfall-induced landslide susceptibility map is derived, and landslide points are classified into five susceptive types. The experimental results show that the BPNN model achieved the best performance. The accuracy of the models is validated using the area under the receiver operating characteristic curve (ROC), area under the curve (AUC), accuracy (ACC), and kappa coefficient. The results showed that all models performed well, but the BPNN model achieved the best performance with an AUC of 0.75, ACC of 0.67, and kappa coefficient of 0.75.

Nowadays, Chinese power grid has developed very rapidly, and the transmission lines are massive. Our paper describes the use of an adaptive dynamic threshold algorithm and machine learning methods to detect wildfires in Yunnan province using Himawari-8. The algorithm extracts relevant features from the original NetCDF images and uses a dynamic threshold to identify wildfire pixels based on solar zenith angle and the proportion of cloud and non-vegetation pixels. Machine learning classifiers, including FCM+ SMOTE+SVM, are trained on the data using techniques to balance the dataset due to data imbalance. The improved classifier performs the best with a high accuracy for fire and non-fire pixels, outperforming other approaches including adaptive dynamic threshold, isolated forest, and one-class support vector machines. The FCM+SMOTE+SVM approach is shown to be robust for wildfire detection, but more data is needed to further improve its performance.

Landslide is one of the major geological disasters in China, which brings huge economic losses to our people every year. However, in the field of landslide treatment, the application of machine learning is scarce. In order to fill the gap in the field of landslide treatment measures based on machine learning. Firstly, random forest classification or regression algorithm was used to train and forecast each landslide treatment measure in this paper. Accuracy (ACC) was used to test the model accuracy of classification algorithm, and Mean Absolute Error (MAE) is used to test the model accuracy of regression algorithm. Random forest classification algorithm was adopted for non-numerical measures. And random forest regression algorithm was adopted for the numerical treatment measures. Secondly, the feature importance of the random forest model was calculated to obtain the more important features of each landslide treatment measure in this paper. Based on this, an optimized random forest model was constructed, and finally the optimal random forest regression and classification algorithm model suitable for landslide treatment measures recommendation was obtained. The training data dimensions of the model were reduced from 58 dimensions to 4-10 dimensions. The experimental results showed that our model could greatly improve the accuracy.

Wildfire is a serious natural disaster that poses a serious threat to the safety of human life and property. Currently, there are many researches related to satellite wildfire detection, but few can achieve near real-time monitoring results. Himawari-8 geostationary satellite can provide full disk data every 10 minutes, making near real-time monitoring of wildfires possible. In this paper, a wildfire detection method based on Himawari-8 for multi-temporal data is proposed. In our method, we use temporal convolutional network (TCN) to predict the brightness temperature and achieve excellent prediction results, the mean absolute error (MAE) is 0.28 K, mean square error (MSE) is 0.30 K2, and mean absolute percentage error (MAPE) is 0.10 %. Then, the predicted values combined with other features as model inputs, and machine learning classification models were used for wildfire detection. The experimental results showed that the combination of multi-layer perceptron (MLP) model and strategy 2 containing brightness temperature predicted values achieved an accuracy of 90.91% in wildfire detection.

Adaptive inference has been proven to improve bidirectional encoder representations from transformers (BERT)'s inference speed with minimal loss of accuracy. However, current work only focuses on the BERT model and lacks exploration of other pretrained language models (PLMs). Therefore, this article conducts an empirical study on the application of adaptive inference mechanism in various PLMs, including generative pretraining (GPT), GCNN, ALBERT, and TinyBERT. This mechanism is verified on both English and Chinese benchmarks, and experimental results demonstrated that it is able to speed up by a wide range from 1 to 10 times if given different speed thresholds. In addition, its application on ALBERT shows that adaptive inference can work with parameter sharing, achieving model compression and acceleration simultaneously, while the application on TinyBERT proves that it can further accelerate the distilled small model. As for the problem that too many labels make adaptive inference invalid, this article also proposes a solution, namely label reduction. Finally, this article open-sources an easy-to-use toolkit called FastPLM to help developers adopt pretrained models with adaptive inference capabilities in their applications.

Facial analysis is an important domain in computer vision and has received extensive research attention. For numerous downstream tasks with different input/output formats and modalities, existing methods usually design task-specific architectures and train them using face datasets collected in the particular task domain. In this work, we proposed a single model, Talk2Face, to simultaneously tackle a large number of face generation and analysis tasks, e.g. text guided face synthesis, face captioning and age estimation. Specifically, we cast different tasks into a sequence-to-sequence format with the same architecture, parameters and objectives. While text and facial images are tokenized to sequences, the annotation labels of faces for different tasks are also converted to natural languages for unified representation. We collect a set of 2.3M face-text pairs from available datasets across different tasks, to train the proposed model. Uniform templates are then designed to enable the model to perform different downstream tasks, according to the task context and target. Experiments on different tasks show that our model achieves better face generation and caption performances than SOTA approaches. On age estimation and multi-attribute classification, our model reaches competitive performance with those models specially designed and trained for these particular tasks. In practice, our model is much easier to be deployed to different facial analysis related tasks. Code and dataset will be available at https://github.com/ydli-ai/Talk2Face.

User cold-start is a major challenge in building personalized recommender systems. Due to the lack of sufficient interactions, it is difficult to effectively model new users. One of the main solutions is to obtain an initial model through meta-learning (mainly gradient-based methods) and adapt it to new users with a few steps of gradient descent. Although these methods have achieved remarkable performance, they are still far from being usable in real-world applications due to their high-demand data processing, heavy computational burden, and inability to perform effective user-incremental update. In this paper, we propose a d eployable and c ontinuable m eta-learning-based r ecommendation (DCMR) approach, which can achieve fast user-incremental updating with task replay and first-order gradient descent. Specifically, we introduce a dual-constrained task sampler, distillation-based loss functions, and an adaptive controller in this framework to balance the trade-off between stability and plasticity in updating. In summary, DCMR can be updated while serving new users; in other words, it learns continuously and rapidly from a sequential user stream and is able to make recommendations at any time. The extensive experiments conducted on three benchmark datasets illustrate the superiority of our model.

Recently, Transformer has been demonstrating promising performance in many NLP tasks and showing a trend of replacing Recurrent Neural Network (RNN). Meanwhile, less attention is drawn to Convolutional Neural Network (CNN) due to its weak ability in capturing sequential and long-distance dependencies, although it has excellent local feature extraction capability. In this paper, we introduce an Attentive Convolutional Transformer (ACT) that takes the advantages of both Transformer and CNN for efficient text classification. Specifically, we propose a novel attentive convolution mechanism that utilizes the semantic meaning of convolutional filters attentively to transform text from complex word space to a more informative convolutional filter space where important n-grams are captured. ACT is able to capture both local and global dependencies effectively while preserving sequential information. Experiments on various text classification tasks and detailed analyses show that ACT is a lightweight, fast, and effective universal text classifier, outperforming CNNs, RNNs, and attentive models including Transformer.

Pengfei Li, Kezhi Mao, Xuefeng Yang, Qi Li. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Virtual Reality (VR) applications allow a user to explore a scene intuitively through a tracked head-mounted display (HMD). However, in complex scenes, occlusions make scene exploration inefficient, as the user has to navigate around occluders to gain line of sight to potential regions of interest. When a scene region proves to be of no interest, the user has to retrace their path, and such a sequential scene exploration implies significant amounts of wasted navigation. Furthermore, as the virtual world is typically much larger than the tracked physical space hosting the VR application, the intuitive one-to-one mapping between the virtual and real space has to be temporarily suspended for the user to teleport or redirect in order to conform to the physical space constraints. In this paper we introduce a method for improving VR exploration efficiency by automatically constructing a multiperspective visualization that removes occlusions. For each frame, the scene is first rendered conventionally, the z-buffer is analyzed to detect horizontal and vertical depth discontinuities, the discontinuities are used to define disocclusion portals which are 3D scene rectangles for routing rays around occluders, and the disocclusion portals are used to render a multiperpsective image that alleviates occlusions. The user controls the multiperspective disocclusion effect, deploying and retracting it with small head translations. We have quantified the VR exploration efficiency brought by our occlusion removal method in a study where participants searched for a stationary target, and chased a dynamic target. Our method showed an advantage over conventional VR exploration in terms of reducing the navigation distance, the view direction rotation, the number of redirections, and the task completion time. These advantages did not come at the cost of a reduction in depth perception or situational awareness, or of an increase in simulator sickness.

Representation learning of words, also known as word embedding technique, is based on the distributional hypothesis that words with similar semantic meanings have similar context. The selection of context window naturally has an influence on word vectors learned. However, it is found that the word vectors are often very sensitive to the defined context window, and unfortunately there is no unified optimal context window for all words. One impact of this issues is that, under a predefined context window, the semantic meanings of some words may not be well represented by the learned vectors. To alleviate the problem and improve word embeddings, we propose a task-independent fine-tuning framework in this paper. The main idea of the task-independent fine tuning is to integrate multiple word embeddings and lexical semantic resources to fine tune a target word embedding. The effectiveness of the proposed framework is tested by tasks of semantic similarity prediction, analogical reasoning, and sentence completion. Experiments results on six word embeddings and eight datasets show that the proposed fine-tuning framework could significantly improve word embeddings.