The development of continuous glucose monitoring (CGM) devices has facilitated research of data-driven glucose prediction, which is useful in managing type 1 diabetes mellitus (T1DM). Existing research has investigated using historical-value based machine learning approaches to predict short-term glucose excursion and adapting pre-trained models to unseen patients via meta-learning. This study aims to investigate if time-index based modeling could improve the effectiveness and efficiency to the same objective. Different to existing methods that take the past glucose sequence as the input and output the future sequence, the method we apply — referred to as the meta-optimised time-index model (MOTIM) — takes time-index features as the input and predicts the glucose value at that time. We argue that the reduction of modeling complexity from sequence-level to value-level could ease the learning difficulty and increase the efficiency. In our experiments on the OhioT1DM and UMT1DM datasets, the MOTIM achieves comparable performance to the state-of-the-art models (e.g., 1-hour excursion averaged mean absolute percentage error of 10.25 on OhioT1DM and 12.50 on UMT1DM) while consuming a substantially lower cost (e.g., only 1.3% model parameters and more than 10 times faster inference). These promising findings encourage further work on time-index modeling in glucose prediction for T1DM, and to encourage it, we have released our codes at https://github.com/r-cui/MOTIMGluPred under the MIT license.

In this study, the high temporal resolution ionospheric product uqrg provided by Universitat Politècnica de Catalunya (UPC) is downsampled into eight products with different temporal resolutions (15, 30 min, 1, 2, 4, 6, 12, and 24 h) in the Antarctic. Then, the accuracy of different products is analyzed using Global Navigation Satellite System (GNSS) total electron content (TEC) from approximately 40 stations’ data between 2015–2016 and 2022–2023. The results indicate that there is little difference in the accuracy of products with time resolution$\le 60$min, with a difference in root mean square (rms) <0.2 Tecu. Then, as the temporal resolution decreased, the rms of products gradually became larger, from 4.2 to 6.0 Tecu in 2015–2016 and 5.1 to 7.0 Tecu in 2022–2023, respectively. The differential TEC between UPC TEC and GNSS TEC is further used to fit a model for correcting UPC products based on the spherical crown harmonic (SCH) function. The results show that the bias and rms decreased from −1.93 to −0.07 and 4.39 to 2.89 Tecu after the correction, and the correction effect is better in polar day than in polar night. Finally, the polynomial model and the long short-term memory (LSTM) model are used to predict the SCH coefficients, respectively, to obtain the predicted TEC. The results suggest that the polynomial model performs better in short-term prediction, with an accuracy improvement of 20.38% compared with the product before corrected, and the LSTM model performs better in long-term prediction, with an accuracy improvement of 13.3%.

Video question answering (VideoQA) is an essential task in vision-language understanding, which has attracted numerous research attention recently. Nevertheless, existing works mostly achieve promising performances on short videos of duration within 15 seconds. For VideoQA on minute-level long-term videos, those methods are likely to fail because of lacking the ability to deal with noise and redundancy caused by scene changes and multiple actions in the video. Considering the fact that the question often remains concentrated in a short temporal range, we propose to first locate the question to a segment in the video and then infer the answer using the located segment only. Under this scheme, we propose “Locate before Answering” (LocAns), a novel approach that integrates a question localization module and an answer prediction module into an end-to-end model. During the training phase, the available answer label not only serves as the supervision signal of the answer prediction module, but also is used to generate pseudo temporal labels for the question localization module. Moreover, we design a decoupled alternative training strategy to update the two modules separately. In the experiments, LocAns achieves state-of-the-art performance on three modern long-term VideoQA datasets, NExT-QA, ActivityNet-QA, and AGQA. Its qualitative examples show the reliable performance of the question localization.

Video moment retrieval aims at finding the start and end timestamps of a moment (part of a video) described by a given natural language query. Fully supervised methods need complete temporal boundary annotations to achieve promising results, which is costly since the annotator needs to watch the whole moment. Weakly supervised methods only rely on the paired video and query, but the performance is relatively poor. In this paper, we look closer into the annotation process and propose a new paradigm called "glance annotation". This paradigm requires the timestamp of only one single random frame, which we refer to as a "glance", within the temporal boundary of the fully supervised counterpart. We argue this is beneficial because comparing to weak supervision, trivial cost is added yet more potential in performance is provided. Under the glance annotation setting, we propose a method named as Video moment retrieval via Glance Annotation (ViGA) based on contrastive learning. ViGA cuts the input video into clips and contrasts between clips and queries, in which glance guided Gaussian distributed weights are assigned to all clips. Our extensive experiments indicate that ViGA achieves better results than the state-of-the-art weakly supervised methods by a large margin, even comparable to fully supervised methods in some cases.

People with type 1 diabetes mellitus (T1DM) must continuously monitor their blood glucose levels and regulate them by insulin dosing to stay in a safe range. A reliable glucose prediction technique could pre-alert the risk of abnormal glycaemia in the near future. In this paper, we apply an attention-based deep network for glucose prediction, which models both the temporal dependencies and the physiological relations among glucose and glucose-related life events, including insulin and carbohydrate intake. We also propose to make use of the knowledge learned from non-target subjects with the same disease to improve the personalised prediction by applying parameters transfer. Our approach was evaluated on the standard benchmark OhioT1DM dataset, where the experiments achieved average root mean square errors over the 12 subjects of 17.82 mg/dL for 30 minutes and 28.54 mg/dL for 60 minutes. Additionally, our ablation experiments indicated that the use of transfer learning constantly improved the prediction. On this basis, we conclude that our approach achieves state-of-the-art performance with statistical significance, and data from other people with T1DM could help on improving personalised predictions. We release our codes at https://github.com/r-cui/GluPred under the MIT license.

Adaptive Cache Replacement Policy (ACRP) has been implemented in recently proposed commercial multi-core processors. ACRP consists of two candidate cache replacement policies and dynamically employs the policy which is with fewer cache misses at the moment. ACRP can diminish the overall cache misses, but at the same time it augments the performance inference between co-running applications and makes the performance prediction much harder. Unfortunately, very little work has focused on the performance impact from this mechanism. In this paper, we firstly expose the performance variation problem due to adaptive cache replacement policies. Secondly, we present Bubble-Bound, a low-overhead measurement-based method to estimate a program's performance variation caused by the dynamic adaptation of cache replacement policies. By using a stress program to characterize the pressure and sensitivity, our method can predict a bound for the performance degradation between co-located applications and enable “safe” co-locations on the processors with ACRP.

Action detection and recognition are popular subjects of research in the field of computer vision. The task of action detection can be regarded as the sum of action location and recognition. Action features described by using information concerning the human skeleton have the advantages of robustness against external factors and requiring a small amount of calculation. This study proposes a skeleton‐based action analysis model based on a recurrent neural network framework. The model learns action features by modelling static and dynamic features of skeleton joints and the importance of different video frames by introducing an attention module. For action location, conditional random field loss function is introduced to establish the context dependency of output labels. In the aspect of action recognition, the hierarchical training mechanism with triple loss models action features at coarse‐grained and fine‐grained levels. The authors’ proposed method delivers state‐of‐the‐art results on action location and recognition tasks.

As the shared last level cache (LLC) in multicore processors has been shown to be a critical resource for system performance, much work has been proposed for improving the quality of service (QoS) and throughput on LLC. Cache Allocation Technology (CAT) and Adaptive Cache Replacement Policies (ACRP) are two of the techniques that are featured in recent Intel processors. CAT implements way partitioning and provides the ability to control the cache space allocation among cores. ACRP works with multiple replacement policies and enables the cache to adapt to the cache replacement policy with less cache misses. In this paper, we first show an interesting finding that ACRP technique can violate the performance isolation provided by CAT. We find the cause for this problem is that the ACRP chooses the cache replacement policy upon the global information even although the cache space partitioning is being enabled by CAT. As the result, the cache/performance isolation can be impaired by the interference on cache replacement policy. To deal with this problem, we propose a low overhead method to predict the worst execution time degradation caused by the replacement policy adaptation. Thus, in the partitioned cache space, if the worst execution time estimated by our method is not beyond the response time required for this program, the QoS for this program can be quaranteed no matter how the cache replacement policies alternate.

Biased decisions made by automatic systems have led to growing concerns in research communities. Recent work from the NLP community focuses on building systems that make fair decisions based on text. Instead of relying on unknown decision systems or human decision-makers, we argue that a better way to protect data providers is to remove the trails of sensitive information before publishing the data. In light of this, we propose a new privacy-aware text rewriting task and explore two privacy-aware back-translation methods for the task, based on adversarial training and approximate fairness risk. Our extensive experiments on three real-world datasets with varying demo-graphical attributes show that our methods are effective in obfuscating sensitive attributes. We have also observed that the fairness risk method retains better semantics and fluency, while the adversarial training method tends to leak less sensitive information.

Skeleton-based action recognition is widely concerned because skeletal information of human body can express action features simply and clearly, and it is not affected by physical features of the human body. Therefore, in this paper, the method of action recognition is based on skeletal information extracted from RGBD video. Since the skeleton coordinates we studied are two-dimensional, our method can be applied to RGB video directly. The recently proposed method based on the deep network only focuses on the temporal dynamic of action and ignores spatial configuration. In this paper, a Multi-source model is proposed based on the fusion of the temporal and spatial models. The temporal model is divided into three branches, which perceive the global-level, local-level, and detail-level information respectively. The spatial model is used to perceive the relative position information of skeleton joints. The fusion of the two models is beneficial to improve the recognition accuracy. The proposed method is compared with the state-of-the-art methods on a large scale dataset. The experimental results demonstrate the effectiveness of our method.