Dual Memory Research Articles

A robust multi-scale feature extraction framework with dual memory module for multivariate time series anomaly detection

Although existing reconstruction-based multivariate time series anomaly detection (MTSAD) methods have shown advanced performance, most assume the training data is clean. When faced with noise or contamination in training data, they can also reconstruct the anomaly well, weakening the distinction between normal and anomaly. Some probabilistic generation-based methods have been used to address this issue because of their implicit robust structure to noise, but the training process and suppression of anomalous generalization are not stable. The recently proposed explicit method based on the memory module would also sacrifice the reconstruction effect of normal patterns, resulting in limited performance improvement. Moreover, most existing MTSAD methods use a single fixed-length window for input, which weakens their ability to extract long-term dependency. This paper proposes a robust multi-scale feature extraction framework with the dual memory module to comprehensively extract features fusing different levels of semantic information and lengths of temporal dependency. First, this paper designs consecutive neighboring windows as inputs to allow the model to extract local and long-term dependency information. Secondly, a dual memory-augmented encoder is proposed to extract global typical patterns and local common features. It ensures the reconstruction ability of normal data while suppressing the generalization of the anomaly. Finally, this paper proposes a multi-scale fusion module to fuse latent variables representing different levels of semantic information and uses the reconstructed latent variables to reconstruct samples for anomaly detection. Experimental results on five datasets from diverse domains show that the proposed method outperforms 16 typical baseline methods.

Detecting turning points of stock markets in China and the United States

ABSTRACT Choosing the appropriate method to identify turning points is critical for investors and policymakers. We propose a novel hybrid model combining dual long memory with structural breaks in the mean to identify the turning points of the stock market in China and the United States. The results show that models accounting for long memory and structural break perform well in the turning point detection. Notably, the proposed hybrid model generates superior in-sample turning point matching and out-sample forecasts for SSECI over those obtained from the competing ARFIMA models. In contrast, DJIA is better explained by the dual memory model with structural breaks in volatility. Finally, our analysis extends to encompass FTSE, HSI, IBOVESPA, and SENSEX30. The results demonstrate the superior performance of the hybrid model in markets exhibiting fractal characteristics. Such heterogeneity indicates differences in investor risk preference across stock markets and offers elaborate international evidence to the adaptive market hypothesis.

A Dual Memory Hybrid Neural Networks for Modeling and Prediction of Nonlinear Systems

A Dual Memory Hybrid Neural Networks for Modeling and Prediction of Nonlinear Systems

Dual memory model for experience-once task-incremental lifelong learning

Experience replay (ER) is a widely-adopted neuroscience-inspired method to perform lifelong learning. Nonetheless, existing ER-based approaches consider very coarse memory modules with simple memory and rehearsal mechanisms that cannot fully exploit the potential of memory replay. Evidence from neuroscience has provided fine-grained memory and rehearsal mechanisms, such as the dual-store memory system consisting of PFC-HC circuits. However, the computational abstraction of these processes is still very challenging. To address these problems, we introduce the Dual-Memory (Dual-MEM) model emulating the memorization, consolidation, and rehearsal process in the PFC-HC dual-store memory circuit. Dual-MEM maintains an incrementally updated short-term memory to benefit current-task learning. At the end of the current task, short-term memories will be consolidated into long-term ones for future rehearsal to alleviate forgetting. For the Dual-MEM optimization, we propose two learning policies that emulate different memory retrieval strategies: Direct Retrieval Learning and Mixup Retrieval Learning. Extensive evaluations on eight benchmarks demonstrate that Dual-MEM delivers compelling performance while maintaining high learning and memory utilization efficiencies under the challenging experience-once setting.

Dual Memory Aggregation Network for Event-Based Object Detection with Learnable Representation

Event-based cameras are bio-inspired sensors that capture brightness change of every pixel in an asynchronous manner. Compared with frame-based sensors, event cameras have microsecond-level latency and high dynamic range, hence showing great potential for object detection under high-speed motion and poor illumination conditions. Due to sparsity and asynchronism nature with event streams, most of existing approaches resort to hand-crafted methods to convert event data into 2D grid representation. However, they are sub-optimal in aggregating information from event stream for object detection. In this work, we propose to learn an event representation optimized for event-based object detection. Specifically, event streams are divided into grids in the x-y-t coordinates for both positive and negative polarity, producing a set of pillars as 3D tensor representation. To fully exploit information with event streams to detect objects, a dual-memory aggregation network (DMANet) is proposed to leverage both long and short memory along event streams to aggregate effective information for object detection. Long memory is encoded in the hidden state of adaptive convLSTMs while short memory is modeled by computing spatial-temporal correlation between event pillars at neighboring time intervals. Extensive experiments on the recently released event-based automotive detection dataset demonstrate the effectiveness of the proposed method.

Dual Memory Units with Uncertainty Regulation for Weakly Supervised Video Anomaly Detection

Learning discriminative features for effectively separating abnormal events from normality is crucial for weakly supervised video anomaly detection (WS-VAD) tasks. Existing approaches, both video and segment level label oriented, mainly focus on extracting representations for anomaly data while neglecting the implication of normal data. We observe that such a scheme is sub-optimal, i.e., for better distinguishing anomaly one needs to understand what is a normal state, and may yield a higher false alarm rate. To address this issue, we propose an Uncertainty Regulated Dual Memory Units (UR-DMU) model to learn both the representations of normal data and discriminative features of abnormal data. To be specific, inspired by the traditional global and local structure on graph convolutional networks, we introduce a Global and Local Multi-Head Self Attention (GL-MHSA) module for the Transformer network to obtain more expressive embeddings for capturing associations in videos. Then, we use two memory banks, one additional abnormal memory for tackling hard samples, to store and separate abnormal and normal prototypes and maximize the margins between the two representations. Finally, we propose an uncertainty learning scheme to learn the normal data latent space, that is robust to noise from camera switching, object changing, scene transforming, etc. Extensive experiments on XD-Violence and UCF-Crime datasets demonstrate that our method outperforms the state-of-the-art methods by a sizable margin.

Experience Replay Optimisation via ATSC and TSC for Performance Stability in Deep RL

Catastrophic forgetting is a significant challenge in deep reinforcement learning (RL). To address this problem, researchers introduce the experience replay (ER) concept to complement the training of a deep RL agent. However, the buffer size, experience selection, and experience retention strategies adopted for the ER can negatively affect the agent’s performance stability, especially for complex continuous state action problems. This paper investigates how to address the stability problem using an enhanced ER method that combines a replay policy network, a dual memory, and an alternating transition selection control (ATSC) mechanism. Two frameworks were designed: an experience replay optimisation via alternating transition selection control (ERO-ATSC) without a transition storage control (TSC) and an ERO-ATSC with a TSC. The first is a hybrid of experience replay optimisation (ERO) and dual-memory experience replay (DER) and the second, which has two versions of its kind, integrates a transition storage control (TSC) into the first framework. After comprehensive experimental evaluations of the frameworks on the pendulum-v0 environment and across multiple buffer sizes, retention strategies, and sampling ratios, the reward version of ERO-ATSC with a TSC exhibits superior performance over the first framework and other novel methods, such as the deep deterministic policy gradient (DDPG) and ERO.

Mitigating Catastrophic Forgetting with Complementary Layered Learning

Catastrophic forgetting is a stability–plasticity imbalance that causes a machine learner to lose previously gained knowledge that is critical for performing a task. The imbalance occurs in transfer learning, negatively affecting the learner’s performance, particularly in neural networks and layered learning. This work proposes a complementary learning technique that introduces long- and short-term memory to layered learning to reduce the negative effects of catastrophic forgetting. In particular, this work proposes the dual memory system in the non-neural network approaches of evolutionary computation and Q-learning instances of layered learning because these techniques are used to develop decision-making capabilities for physical robots. Experiments evaluate the new learning augmentation in a multi-agent system simulation, where autonomous unmanned aerial vehicles learn to collaborate and maneuver to survey an area effectively. Through these direct-policy and value-based learning experiments, the proposed complementary layered learning is demonstrated to significantly improve task performance over standard layered learning, successfully balancing stability and plasticity.

A novel memory-based simulated annealing algorithm to solve multi-line facility layout problem

In this paper, a memory-based simulated annealing algorithm called the Dual Memory Simulated Annealing Algorithm (DMSA) is presented to solve multi-line facility layout problems. The objective is to minimize the total material handling cost. Two memory buffers and a restart mechanism are considered. Two benchmark problems were selected from the literature review papers and solved using the standard simulated annealing (SA) algorithm and the DMSA. The obtained results show that solutions provided by the DMSA algorithm are cost-effective compared to the standard SA algorithm and the other algorithms used for solving these test cases. Moreover, to further evaluate the performance of the DMSA algorithm in large scale problems, eleven test cases were selected from the benchmark library of the quadratic assignment problem (QAP). According to the results, the performance of the algorithm in finding solutions to complex problems is exemplary.

Age does not modify the processing architecture of dual memory retrieval: an investigation of age-related effects on dual-retrieval practice in younger and older adults

ABSTRACT The present study investigated the cognitive processing architecture of dual(-memory) retrieval from a single cue across two distinct age groups: younger and older adults. Previous research has shown that younger adults can exhibit learned parallel retrieval, but only if they synchronize response execution. This phenomenon has not been demonstrated with older adults. Experiment 1 functioned as an extension of previous studies to assess whether the finding of learned retrieval parallelism in younger adults could be observed in older adults as well. The experiment used a dual retrieval task that involved the retrieval of two responses, one vocal and one keypress, from a single cue. Experiment 2 further assessed whether the cognitive processing architecture underlying the occurrence of learned retrieval parallelism in dual memory retrieval could be influenced by the number of cues in single-retrieval practice. The results of both experiments showed that learned retrieval parallelism occurs in older as well as younger adults and that the processing mechanisms involved in dual memory retrieval are relatively stable across age groups.

Three-Dimensional Path Planning for Unmanned Helicopter Using Memory-Enhanced Dueling Deep Q Network

Unmanned helicopter (UH) is often utilized for raid missions because it can evade radar detection by flying at ultra-low altitudes. Path planning is the key technology to realizing the autonomous action of UH. On the one hand, the dynamically changing radar coverage area and the mountains in the low airspace environment will seriously affect the flight safety of UH. On the other hand, the huge state space of the three-dimensional (3D) environment will also make traditional algorithms difficult to converge. To address the above problems, a memory-enhanced dueling deep Q-network (ME-dueling DQN) algorithm was proposed. First, a comprehensive reward function was designed, which can guide the algorithm to converge quickly and effectively improve the sparse reward problem. Then, we introduced a dual memory pool structure and proposed a memory-enhanced mechanism, which can reduce invalid exploration, further improve the learning efficiency of the algorithm, and make the algorithm more stable. Finally, the path planning ability of the proposed algorithm in multiple experimental environments was verified. Experiments showed that the proposed algorithm has good environmental adaptability and can help UH to accurately identify dangerous areas and plan a safe and reliable flight path.

Dual memory scale network for multi-step time series forecasting in thermal environment of aquaculture facility: A case study of recirculating aquaculture water temperature

Multi-step time series forecasting is essential in engineering. However, effective time series prediction of the agricultural environment is still a challenge due to the disturbance of external factors and the complexity of temporal patterns. The objective of this work is to provide an intelligent prediction system with field priori-knowledge for aquaculture facility thermal environment. Specifically, we present a novel network framework that considers the spatial correlation of exogenous environmental factors, the short-term and long-term temporal dependence of the sequence, and the spatio-temporal fusion correlation at different times. Among them, the dual-scale memory components focus on long-term seasonal and trend memory patterns as well as short-term fluctuation memory patterns, respectively. Besides, the attention mechanism coordinates the information flow between the long and short memory components. We applied the model to recirculating water temperature forecasting in the aquaponics greenhouse. The empirical results of global and local metrics indicate that our proposed model outperforms benchmark models. The RMSE of the proposed dual memory scale network (DMSNet) is 0.1631, 0.3206, and 0.3761, respectively, in the next 6 h, 12 h, and 24 h water temperature prediction. Ablation study proved that the components of our proposed network could be targeted to extract different information patterns from aquaponics environmental sequence data. Experimental results show that the thermal environment prediction based on DMSNet has higher precision in short-term and long-term forecasting multi-step prediction.

Lifelong 3D object recognition and grasp synthesis using dual memory recurrent self-organization networks

Humans learn to recognize and manipulate new objects in lifelong settings without forgetting the previously gained knowledge under non-stationary and sequential conditions. In autonomous systems, the agents also need to mitigate similar behaviour to continually learn the new object categories and adapt to new environments. In most conventional deep neural networks, this is not possible due to the problem of catastrophic forgetting, where the newly gained knowledge overwrites existing representations. Furthermore, most state-of-the-art models excel either in recognizing the objects or in grasp prediction, while both tasks use visual input. The combined architecture to tackle both tasks is very limited. In this paper, we proposed a hybrid model architecture consists of a dynamically growing dual-memory recurrent neural network (GDM) and an autoencoder to tackle object recognition and grasping simultaneously. The autoencoder network is responsible to extract a compact representation for a given object, which serves as input for the GDM learning, and is responsible to predict pixel-wise antipodal grasp configurations. The GDM part is designed to recognize the object in both instances and categories levels. We address the problem of catastrophic forgetting using the intrinsic memory replay, where the episodic memory periodically replays the neural activation trajectories in the absence of external sensory information. To extensively evaluate the proposed model in a lifelong setting, we generate a synthetic dataset due to lack of sequential 3D objects dataset. Experiment results demonstrated that the proposed model can learn both object representation and grasping simultaneously in continual learning scenarios.

Simplified and Detailed Analysis of Data Center Particulate Contamination at Server and Room Level Using Computational Fluid Dynamics

Abstract Continuous rise in cloud computing and other web-based services propelled the data center proliferation seen over the past decade. Traditional data centers use vapor-compression-based cooling units that not only reduce energy efficiency but also increase operational and initial investment costs due to involved redundancies. Free air cooling and airside economization can substantially reduce the information technology equipment (ITE) cooling power consumption, which accounts for approximately 40% of energy consumption for a typical air-cooled data center. However, this cooling approach entails an inherent risk of exposing the ITE to harmful ultrafine particulate contaminants, thus, potentially reducing the equipment and component reliability. The present investigation attempts to quantify the effects of particulate contamination inside the data center equipment and ITE room using computational fluid dynamics (CFD). An analysis of the boundary conditions to be used was done by detailed modeling of ITE and the data center white space. Both two-dimensional and three-dimensional simulations were done for detailed analysis of particle transport within the server enclosure. An analysis of the effect of the primary pressure loss obstructions like heat sinks and dual inline memory modules inside the server was done to visualize the localized particle concentrations within the server. A room-level simulation was then conducted to identify the most vulnerable locations of particle concentration within the data center space. The results show that parameters such as higher velocities, heat sink cutouts, and higher aspect ratio features within the server tend to increase the particle concentration inside the servers.

Multivariate FMRI Signatures of Learning in a Hebb Repetition Paradigm With Tone Sequences.

Important information from the environment often arrives to the brain in temporally extended sequences. Language, music, actions, and complex events generally unfold over time. When such informational sequences exceed the limited capacity of working memory, the human brain relies on its ability to accumulate information in long-term memory over several encounters with a complex stimulus. A longstanding question in psychology and neuroscience is whether the neural structures associated with working memory storage—often viewed as capacity limited and temporary—have any builtin ability to store information across longer temporal delays. According to the classic Hebbian dual memory theory, temporally local “activity traces” underlie immediate perception and working memory, whereas “structural traces” undergird long-term learning. Here we examine whether brain structures known to be involved in working maintenance of auditory sequences, such as area Spt, also show evidence of memory persistence across trials. We used representational similarity analysis (RSA) and the Hebb repetition paradigm with supracapacity tonal sequences to test whether repeated sequences have distinguishable multivoxel activity patterns in the auditory-motor networks of the brain. We found that, indeed, area Spt and other nodes of the auditory dorsal stream show multivoxel patterns for tone sequences that become gradually more distinct with repetition during working memory for supracapacity tone-sequences. The findings suggest that the structures are important for working memory are not “blank slates,” wiped clean from moment to moment, but rather encode information in a way can lead to cross-trial persistence.

Incremental human action recognition with dual memory

Incremental learning is a topic of great interest in the current state of machine learning research. Real-world problems often require a classifier to incorporate new knowledge while preserving what was learned before. One of the most challenging problems in computer vision is Human Action Recognition (HAR) in videos. However, most of the existing works approach HAR from a non-incremental point of view. This work proposes a framework for performing HAR in the incremental learning scenario called Incremental Human Action Recognition with Dual Memory (IHAR-DM). IHAR-DM contains three main components: a 3D convolutional neural network for capturing Spatio-temporal features; a Triplet Network to perform metric learning; and the dual-memory Extreme Value Machine, which is introduced in this work. The proposed method is compared with 10 other state-of-the-art incremental learning models. We propose five experimental settings containing different numbers of tasks and classes using two widely known HAR datasets: UCF-101 and HMDB51. Our results show superior performance in terms of Normalized Mutual Information (NMI) and Inter-task Intransigence (ITI), which is a new metric proposed in this work. Overall results show the feasibility of the proposal for real HAR problems, which mostly present the requirements imposed by incremental learning.

Design and implementation of Dual-Port Memory

Multiport memory cell using a dual-port memory cell provides required access to multi-processor-based applications. Simultaneous access can be provided using two-pass transistors, pair of bit lines, and a word line. Using specific word lines and bit lines of SRAM cell access can be provided by using dual ports memory. The single address of a memory cell can be accessed at a time during each clock pulse using single-port SRAM this drawback can be overcome by using dual-port RAM which supports concurrent read or writes access at different addresses. Efficiency is improved by using dual-port RAM. Each processor can be made to operate at different clock frequencies thereby dual-port RAM will not have any limitations of access between the two ports.

Dual Memory LSTM with Dual Attention Neural Network for Spatiotemporal Prediction.

Spatiotemporal prediction is challenging due to extracting representations being inefficient and the lack of rich contextual dependences. A novel approach is proposed for spatiotemporal prediction using a dual memory LSTM with dual attention neural network (DMANet). A new dual memory LSTM (DMLSTM) unit is proposed to extract the representations by leveraging differencing operations between the consecutive images and adopting dual memory transition mechanism. To make full use of historical representations, a dual attention mechanism is designed to capture long-term spatiotemporal dependences by computing the correlations between the current hidden representations and the historical hidden representations from temporal and spatial dimensions, respectively. Then, the dual attention is embedded into DMLSTM unit to construct a DMANet, which enables the model with greater modeling power for short-term dynamics and long-term contextual representations. An apparent resistivity map (AR Map) dataset is proposed in this paper. The B-spline interpolation method is utilized to enhance AR Map dataset and makes apparent resistivity trend curve continuous derivative in the time dimension. The experimental results demonstrate that the developed method has excellent prediction performance by comparisons with some state-of-the-art methods.

Thermo-Responsive Shape Memory Behavior of Methyl Vinyl Silicone Rubber/Olefin Block Copolymer Blends via Co-Crosslinking

Abstract Shape memory polymers (SMPs) are developed by blending and cross-linking polymers which include crystalline domains and cross-linked networks. In this paper, we describe the morphology, thermal and shape memory behavior of methyl vinyl silicone rubber (MVMQ)/olefin block copolymer (OBC) blends prepared by a melt-blending and chemical cross-linking method. MVMQ without crystalline domains could not hold its temporary shape. After introducing the OBC, the obtained blends exhibited excellent dual shape memory properties. The cross-linking networks of MVMQ acted as reversible domains, while crystalline regions of OBC worked as fixed domains. When the blending ratio of MVMQ/OBC was 50/ 50, the blend had both a high shape fixity ratio and shape recovery ratio.

Deep industrial transfer learning at runtime for image recognition

AbstractThe utilization of deep learning in the field of industrial automation is hindered by two factors: The amount and diversity of training data needed as well as the need to continuously retrain as the use case changes over time. Both problems can be addressed by industrial deep transfer learning allowing for the performant, continuous and potentially distributed training on small, dispersed datasets. As a specific example, a dual memory algorithm for computer vision problems is developed and evaluated. It shows the potential for state-of-the-art performance while being trained only on fractions of the complete ImageNet dataset at multiple locations at once.