Language models face an inherent challenge in balancing memory capacity with the need to efficiently discard irrelevant patterns. We propose a novel adaptive layer-targeted regularisation framework that enables controlled forgetting by integrating temporal decay scheduling with exponential regularisation strength reduction, employing layer-specific targeting on final multi-layer perceptron (MLP) components, and applying L2 norm constraints selectively to memorisation-critical parameters. Our experiments on a 6-layer transformer reveal that targeted L2 regularisation on final-layer MLP weights maintains validation performance while improving inference speed by 19.2%. Validation on enwik8 dataset using generative pre-trained transformer 2 (GPT2)-small architecture demonstrates 6.58% speedup with only 0.52% validation loss increase, confirming broad applicability. We demonstrate that later transformer layers disproportionately influence memorisation patterns (containing 83% of memorisation capacity), enabling precise interventions without compromising linguistic capabilities. Our approach creates a predictable efficiency-performance trade-off with minimal training overhead. This work offers immediate practical benefits for reducing inference costs and energy consumption in production language models.
The existing aspect-based sentiment analysis methods have achieved significant success. However, they face the challenge of low accuracy in handling multiple-aspect words due to long-range dependencies and insufficient contextual information which results in interference from irrelevant words. A hybrid attention network and sentiment lexicon integrated model (HAN-SL) is proposed to solve the problem. The model introduces a phrase structure tree to segment long sentences, minimising the interference of irrelevant words, and integrating a sentiment lexicon to generate a sentiment enhancement matrix that enriches contextual information. In addition, the HAN-SL constructs a matrix between aspect words and viewpoint words, capturing intrinsic grammatical features through aspect and context-aware attention. The fusion of a sentiment enhancement matrix with a sentiment lexicon effectively enriches contextual semantic information. A dual-channel GCN extracts representation information from two matrices to facilitate efficient information exchange. Extensive experimental results on four public datasets demonstrate the superiority and rationality of HAN-SL.
The growing complexity of modern networks due to diverse internet services challenges efficient routing and load balancing. Traditional single-path routing fails to handle such dynamic requirements effectively. This paper proposes an enhanced traffic-aware multipath routing (TAMR) scheme in software-defined networks (SDNs), which leverages the centralised programmability of SDNs to dynamically measure real-time network bandwidth and select multiple optimal paths using Yen's algorithm. The proposed TAMR scheme achieves superior bandwidth utilisation, reduced transmission delay, and improved load balancing under varying traffic conditions by integrating a packet monitoring module to manage out-of-order packets. The innovation lies in its adaptive multipath routing framework, which combines real-time network awareness and packet monitoring to ensure reliable data delivery. The results demonstrate the industrial applicability of TAMR in datacentre networks and high-demand environments requiring low latency and high throughput, highlighting its potential for enhancing network performance and resilience.
Multimodal aspect-based sentiment analysis, as a fine-grained sentiment analysis task, aims to identify the sentiment polarity associated with a given aspect entity. Current methods mainly focus on fusing information from different sources (e.g., image and text). However, in real-world scenarios, the emotional contributions of different modalities are not balanced. Therefore, directly exploring modality fusion inevitably has its limitations. To address this issue, we propose a multimodal aspect-based sentiment analysis network with adaptive modality balancing (AMB), which adaptively analyses cross-modal emotional contributions. Specifically, we first separately aggregate the feature information of a single modality through self-attention. Furthermore, during modality fusion, each modality is assigned a different weight based on its feature entropy, which helps balance their emotional contributions. Experiments on two public datasets validate the effectiveness of the proposed network.
Mental health disorders affect over 25% of the global population, making scalable detection essential for early intervention. Current social media datasets often use community-assigned or platform-inferred labels, which may lack semantic clarity and category consistency. This study explores whether large language models (LLMs), such as GPT-4, can generate more reliable annotations by leveraging prompt-based reasoning aligned with standardised symptom criteria. Using a Reddit dataset of 17,159 posts, we re-annotate entries using a chain-of-thought (CoT) framework guided by symptom profiles from screening instruments. We then evaluate these LLM-generated annotations against subreddit-derived labels via two downstream tasks: 1) classification performance under supervised learning; 2) clustering under unsupervised methods. Results show that LLM-generated annotations yield higher consistency and improve downstream performance, particularly for depression and anxiety, demonstrating their potential to enhance mental health detection from online text.
Epilepsy is a widespread neurological condition impacting people of all ages. Medical professionals use electroencephalography (EEG) as a monitoring tool to analyse neural activity and detect signs of epilepsy. This issue often stems from the inclusion of superfluous EEG features - such as noise and irrelevant data - that fail to support accurate diagnosis. Therefore, feature selection (FS) methods are necessary to filter out irrelevant features and retain the most diagnostically significant ones. This study proposes an innovative FS method using the fossa optimisation algorithm (FSFOA) to determine the most effective feature subset for improving classification accuracy. This method is compared with four widely recognised FS techniques: ACO, GA, PSO, and DE. The evaluation is conducted using five popular classifiers: QDA, NB, DT, SVM and KNN. Experimental results reveal that the proposed FSFOA outperforms the aforementioned methods in selecting optimal features and enhancing classification performance.
Increasing adoption of internet of things (IoT) technologies in agriculture necessitates reliable and scalable sensor networks for smart farming applications. This work proposes the design and evaluation of a relay-based multiplexer for IoT-enabled agricultural systems enhancing sensor interconnectivity while minimising input/output port requirements. To address limited connectivity and resource constraints commonly encountered in agricultural environments, the system employs cost-effective magnetic relay technology. Strategically placed relays enable seamless integration of a large number of spatially distributed sensors across extensive and challenging terrains. Adaptive sensor integration algorithms dynamically optimise data transmission channels, resulting in reduced energy consumption and improved network efficiency. Experimental results demonstrate that the proposed multiplexer can connect and manage a significantly higher number of sensors, leading to enhanced data accuracy, real-time monitoring, and informed decision-making for precision agriculture. The system supports the connection of S & times; N homogeneous sensors, where S represents select lines and N denotes analog ports, enabling sensor scalability beyond the available input pins.
In the initialisation process of monocular visual SLAM, environmental factors, lighting conditions, and moving objects can cause errors in feature matching. We propose an optimised method for monocular visual localisation based on an improved version of the oriented fast and rotated brief (ORB) algorithm. By replacing rBRIEF descriptors with box average difference (BAD) descriptors, we enhance the robustness and computational efficiency of the feature matching algorithm. The use of the PROSAC algorithm eliminates mismatches and further improves accuracy. Experimental results show that our method improves the accuracy of feature matching in complex environments while maintaining computational efficiency. Compared to ORB, our method reduces runtime by approximately 10 ms. In typical complex scenes, the reprojection error is close to the sub-pixel level, and the matching accuracy is improved by 1.31% to 15.32%. Our method enhances computational efficiency during the initialisation process of monocular visual SLAM, thereby indirectly improving localisation accuracy. It is applicable in fields such as autonomous driving, robot navigation, mining exploration, virtual reality, and more.
