Temporal Memory Network Research Articles

Clustered temporal memory networks: A hybrid approach for signal strength prediction

The rapid expansion of 5G networks has revolutionized mobile communication by offering unprecedented speed, low-latency connections, and the ability to support vast numbers of connected devices. However, these advancements bring new challenges in maintaining consistent and reliable signal strength, critical for ensuring optimal Quality of Service (QoS). Traditional models, such as ARIMA, Random Forest (RF), and K-means clustering, struggle to capture the complex, nonlinear, and dynamic behaviour of 5G networks, leading to suboptimal prediction accuracy. In this study, we propose a novel hybrid model, Clustered Temporal Memory Networks (CTMN), which integrates DBSCAN clustering with Long Short-Term Memory (LSTM) networks to improve signal strength prediction in mobile networks. The CTMN model combines DBSCAN's ability to handle spatial variability and outliers in 5G data, combined with LSTM's capacity for modelling long-term dependencies and nonlinear time-series patterns. Our empirical analysis demonstrates that CTMN outperforms traditional methods, achieving up to a 20.82% improvement in prediction accuracy across key performance metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). These findings indicate that CTMN provides a scalable, robust solution for enhancing signal strength prediction and optimizing network performance in next-generation mobile networks.

Longitudinal relationships between Aβ and tau to executive function and memory in cognitively normal older adults

The early accumulation of AD pathology such as Aβ and tau in cognitively normal older people is predictive of cognitive decline, but it has been difficult to dissociate the cognitive effects of these two proteins. Early Aβ and tau target distinct brain regions that have different functional roles. Here, we assessed specific longitudinal pathology-cognition associations in seventy-six cognitively normal older adults from the Berkeley Aging Cohort Study who underwent longitudinal PiB PET, FTP PET, and cognitive assessments. Using linear mixed-effects models to estimate longitudinal changes and residual approach to characterizing cognitive domain-specific associations, we found that Aβ accumulation, especially in frontal/parietal regions, was associated with faster decline in executive function, not memory, whereas tau accumulation, especially in left entorhinal/parahippocampal regions, was associated with faster decline in memory, not executive function, supporting an “Aβ-executive function, tau-memory” double-dissociation in cognitively normal older people. These specific relationships between accumulating pathology and domain-specific cognitive decline may be due to the particular vulnerabilities of the frontal-parietal executive network to Aβ and temporal memory network to tau.

A novel spatio-temporal attention-based bidirectional LSTM model for moisture content prediction in drying process

Accurate prediction of moisture content is significantly crucial for ensuring process stability and product quality in the cylinder drying process. However, the drying process exhibits complex spatio-temporal characteristics and strong interference, which make accurate prediction challenging for the deep learning approach. To address this issue, this article proposes a new spatio-temporal attention-based bidirectional long-short temporal memory network (STA-BiLSTM) model for accurate moisture content prediction. First, Maximum Relevance Minimum Redundancy (mRMR) is adopted to identify optimal features highly related to moisture content. Secondly, bidirectional long-short temporal memory (Bi-LSTM) network is utilized to extract temporal dependencies from the sequential data. Subsequently, spatio-temporal attention mechanisms are designed to adaptively focus on the most relevant features and timesteps, enhancing the model’s generalization ability. Finally, due to the harsh industrial environment, eXtreme Gradient Boosting (XGBoost) is adapted to improve generalizability and robustness. Extensive experiments on a real industrial dataset of the drying process demonstrate that the proposed STA-BiLSTM approach significantly outperforms alternative approaches for predicting moisture content, validating its effectiveness and superiority.

Dual temporal memory network with high-order spatio-temporal graph learning for video object segmentation

Typically, Video Object Segmentation (VOS) always has the semi-supervised setting in the testing phase. The VOS aims to track and segment one or several target objects in the following frames in the sequence, only given the ground-truth segmentation mask in the initial frame. A fundamental issue in VOS is how to best utilize the temporal information to improve the accuracy. To address the aforementioned issue, we provide an end-to-end framework that simultaneously extracts long-term and short-term historical sequential information to current frame as temporal memories. The integrated temporal architecture consists of a short-term and a long-term memory modules. Specifically, the short-term memory module leverages a high-order graph-based learning framework to simulate the fine-grained spatial–temporal interactions between local regions across neighboring frames in a video, thereby maintaining the spatio-temporal visual consistency on local regions. Meanwhile, to relieve the occlusion and drift issues, the long-term memory module employs a Simplified Gated Recurrent Unit (S-GRU) to model the long evolutions in a video. Furthermore, we design a novel direction-aware attention module to complementarily augment the object representation for more robust segmentation. Our experiments on three mainstream VOS benchmarks, containing DAVIS 2017, DAVIS 2016, and Youtube-VOS, demonstrate that our proposed solution provides a fair tradeoff performance between both speed and accuracy.

Forecasting call center arrivals using temporal memory networks and gradient boosting algorithm

Forecasting call center arrivals using temporal memory networks and gradient boosting algorithm

Spatio-temporal AI inference engine for estimating hard disk reliability

This paper focuses on building a spatio-temporal AI inference engine for estimating hard disk reliability. Most electronic systems such as hard disks routinely collect such reliability parameters in the field to monitor the health of the system. Changes in parameters as a function of time are monitored and any observed changes are compared with the known failure signatures. If the trajectory of the measured data matches that of a failure signature, operators are alerted to take corrective action. However, the interest of the operators lies in being able to identify the failures before they occur. The state of the art methodology including our prior work is to train machine learning models on temporal sequence data capturing the variations across multiple features and using it to predict the remaining useful life of the devices. However, as we show in this paper temporal prediction capability alone is not sufficient and can lead to low precision and the uncertainty around the prediction is very large. This is primarily due to the non-uniform progression of feature patterns over time. Our hypothesis is that the accuracy can be improved if we combine the temporal prediction methods with a spatial analysis that compares the value of key SMART features of the devices across similar model in a fixed time window (unlike the temporal method which uses the data from a single device and a much larger historical window). In this paper, we first describe both temporal and spatial approaches, describe the methods to select various hyperparameters, and then show a workflow to combine these two methodologies and provide comparative results. Our results illustrate that the average precision of temporal methods using long-short temporal memory networks to predict impending failures in the next ten days was 84 percent. To improve precision, we use the set of disks identified as potential failures and start applying spatial anomaly detection methods on those disks. This helps us remove the false positives from the temporal prediction results and provide a tighter bound on the set of disks with impending failure.

Temporal Memory Network Towards Real-Time Video Understanding

Action recognition is the basic task for video understanding. Although the action recognition has achieved impressive performance in the static image-based task (e.g. Stanford40) with deep learning, real-time video-based action recognition is still a challenging task due to video’s high complexity and computation cost. Motivated by human’s recognition ability with only a short glance, we propose the fast light-weighted Temporal Memory Network (TMNet) to achieve real-time video action recognition. The TMNet has a self-supervised structure for exploring both spatial and temporal information with a single video frame. TMNet has three main parts, the base backbone, the regression branch, and the classification branch. Specifically, the base backbone network is a shallow 2D CNN network to obtain the video’s initial feature sequences. The classification branch is based on existing successful video recognition models(e.g. TSN, I3D). To make the TMNet learn the spatial-temporal information at a lower cost, we add a self-supervised regression branch. This branch is based on light-weighted 2D CNN and only uses one frame as input. In the training stage, the input of TMNet is a video sequence, the classification branch combined with the base backbone is responsible for learning the video sequence’s spatial-temporal feature. Meanwhile, the self-supervised regression branch aims to learn the same spatial-temporal feature under the supervision of the classification branch’s output. And the regression branch’s input is a single-frame feature sampled from the encoded video sequence. In this way, the regression branch is forced to learn temporal information of adjacent frames with one frame. Therefore, TMNet only needs one frame to predict each video’s spatial-temporal information in the inference stage. Finally, TMNet can achieve real-time action recognition and better accuracy by extracting temporal information from a static image. Abundant ablation experiments demonstrate TMNet has a good trade-off between accuracy and speed.

Multi-object recognition by optimized hierarchical temporal memory network

In this paper, hierarchical temporal memory network (HTM) was optimized for multi-object recognition. HTM is constructed by temporal module and spatial module, which is formulated by Hawkins and George in 2005 based on prediction theory. Multi-object recognition is a spatial pattern recognition task, so we reduction the temporal module of hierarchical temporal memory network and strengthen the spatial module. Furthermore, sparse representation method was used for capturing the convolution kernels in the network, which simulates the function of the retina cells of the eyes. Moreover, the prediction space is used in the network to accelerate pattern identification. Finally, a four-level network was designed and trained for locomotive object recognition, and the recognition rate is up to 91.4%.

Verbal and nonverbal memory in primary progressive aphasia: the Three Words-Three Shapes Test.

Objectives: To investigate cognitive components and mechanisms of learning and memory in primary progressive aphasia (PPA) using a simple clinical measure, the Three Words Three Shapes Test (3W3S). Background: PPA patients can complain of memory loss and may perform poorly in standard tests of memory. The extent to which these signs and symptoms reflect dysfunction of the left hemisphere language versus limbic memory network remains unknown. Methods: 3W3S data from 26 patients with a clinical diagnosis of PPA were compared with previously published data from patients with typical dementia of the Alzheimer type (DAT) and cognitively healthy elders. Results: PPA patients showed two bottlenecks in new learning. First, they were impaired in the effortless (but not effortful) on-line encoding of verbal (but not non-verbal) items. Second, they were impaired in the retrieval (but not retention) of verbal (but not non-verbal) items. In contrast, DAT patients had impairments also in effortful on-line encoding and retention of verbal and nonverbal items. Conclusions: PPA selectively interferes with spontaneous on-line encoding and subsequent retrieval of verbal information. This combination may underlie poor memory test performance and is likely to reflect the dysfunction of the left hemisphere language rather than medial temporal memory network.

Verbal and Nonverbal Memory in Primary Progressive Aphasia: The Three Words-Three Shapes Test

Objectives:To investigate cognitive components and mechanisms of learning and memory in primary progressive aphasia (PPA) using a simple clinical measure, the Three Words Three Shapes Test (3W3S).Background:PPA patients can complain of memory loss and may perform poorly in standard tests of memory. The extent to which these signs and symptoms reflect dysfunction of the left hemisphere language versus limbic memory network remains unknown.Methods:3W3S data from 26 patients with a clinical diagnosis of PPA were compared with previously published data from patients with typical dementia of the Alzheimer type (DAT) and cognitively healthy elders.Results:PPA patients showed two bottlenecks in new learning. First, they were impaired in the effortless (but not effortful) on-line encoding of verbal (but not non-verbal) items. Second, they were impaired in the retrieval (but not retention) of verbal (but not non-verbal) items. In contrast, DAT patients had impairments also in effortful on-line encoding and retention of verbal and nonverbal items.Conclusions:PPA selectively interferes with spontaneous on-line encoding and subsequent retrieval of verbal information. This combination may underlie poor memory test performance and is likely to reflect the dysfunction of the left hemisphere language rather than medial temporal memory network.

Gliding and saccadic gaze gesture recognition in real time

Eye movements can be consciously controlled by humans to the extent of performing sequences of predefined movement patterns, or 'gaze gestures'. Gaze gestures can be tracked noninvasively employing a video-based eye tracking system. Gaze gestures hold the potential to become an emerging input paradigm in the context of human-computer interaction (HCI) as low-cost eye trackers become more ubiquitous. The viability of gaze gestures as an innovative way to control a computer rests on how easily they can be assimilated by potential users and also on the ability of machine learning algorithms to discriminate in real time intentional gaze gestures from typical gaze activity performed during standard interaction with electronic devices. In this work, through a set of experiments and user studies, we evaluate the performance of two different gaze gestures modalities, gliding gaze gestures and saccadic gaze gestures, and their corresponding real-time recognition algorithms, Hierarchical Temporal Memory networks and the Needleman-Wunsch algorithm for sequence alignment. Our results show that a specific combination of gaze gesture modality, namely saccadic gaze gestures, and recognition algorithm, Needleman-Wunsch, allows for reliable usage of intentional gaze gestures to interact with a computer with accuracy rates higher than 95% and completion speeds of around 1.5 to 2.5 seconds per gesture. The optimal gaze gesture modality and recognition algorithm do not interfere with otherwise standard human-computer gaze interaction, generating very few false positives during real time recognition and positive feedback from the users. These encouraging results and the low cost eye tracking equipment used, open up a new HCI paradigm for the fields of accessibility and interaction with smartphones, tablets, projected displays and traditional desktop computers.