Specific Tasks Research Articles

A safety realignment framework via subspace-oriented model fusion for large language models

To improve the performance of large language models (LLMs) on specific tasks, task-specific instruction fine-tuning is essential. However, this process can easily compromise the safety of a task-specific model, making it susceptible to obeying malicious instructions and generating harmful content. Current methods against fine-tuning attack usually interfere with the original fine-tuning objectives or require substantial amounts of data to realign the compromised model. To address these two major challenges, we propose reusing the initial aligned model and realigning task-specific model in the safety subspace. In this paper, we introduce a safety realignment framework through subspace-oriented model fusion (SOMF), aiming to transfer the safeguard capabilities of an initially aligned model into the current task-specific model. Our approach begins by disentangling all task vectors from the parameters of each task-specific model. We then identify safety-critical regions within these vectors by subspace masking techniques. Finally, we fuse the initial safely aligned LLM with all task vectors based on the identified safety subspace to restore the model’s safety properties. Our experiments confirm that our safety realignment framework satisfies the safety requirements of an independent task-specific model as well as traditional multitask models during their fusion. Our findings confirm that SOMF preserves safety without notably compromising performance on specific tasks while exhibiting higher data efficiency. The code is publicly available at https://github.com/xinykou/safety_realignment.

Leveraging multimodal deep learning for natural disaster event classification and its damage severity analysis through social media posts

Accurate and timely information is essential for coordinating an effective disaster response. Traditional methods have struggled to efficiently categorize disaster events and assess damage severity due to the variety and complexity of data sources. Previous research has focused on specific tasks, such as information gathering or humanitarian assistance, but has not adequately addressed the assessment of disaster damage severity. This paper proposes a hybrid learning model to improve disaster event classification and damage severity identification. The model combines image and text data in a cooperative way, using ResNet50 to extract features from images and a LSTM with attention mechanism to learn sequences from text. This combination allows for a more contextual and informative representation of the input data. Compared to existing approaches, the proposed multimodal approach achieves significantly better results in disaster event classification. Apart from the proposed model also shows promising outcome for damage severity of disaster. These advancements are especially important for real-world applications such as disaster management and response coordination, where accuracy and reliability are essential. The comprehensive methodology and empirical results presented in this paper demonstrate the effectiveness and potential of using hybrid learning models to leverage multimodal data for unique and sophisticated analytical tasks in disaster scenarios

Transfer graph feature alignment guided multi-source domain adaptation network for machinery fault diagnosis

In recent years, the application of unsupervised multi-source domain adaptation (MSDA) techniques for fault diagnosis has gained significant traction. Current research typically overlooks or fails to effectively capture critical data structure information during feature extraction. Another challenge is optimising the integration of information from multiple source domains to diagnose the target domain while avoiding negative transfer. To address these challenges, this study proposes a transfer graph feature alignment guided multi-source domain adaptation network (MDTGAL). In the proposed method, a transfer graph sample generator module (GSG) is constructed to model the data structure between source and target domains, and multiple graph feature extractors are employed to learn the data structure information from different domain combinations. A regularisation technique is introduced to extract the domain-invariant features by aligning the parameters across multiple independent graph feature extractors. In addition, a weighted soft-voting mechanism based on the polynomial kernel-induced maximum mean difference metric (PK-MMD) is designed to fuse the outputs from multiple classifiers, to comprehensively account for the influence of each source domain. The proposed method was tested on multi-source domain transfer tasks involving various operating conditions of rotating machinery. The experimental results demonstrate that the MDTGAL exhibits superior cross-domain diagnostic performance, outperforming existing mainstream methods. In addition, this study explored the impact of varying numbers of source domains on the diagnostic accuracy of the target domain, providing insights into selecting the correct number of source domains for specific MSDA tasks.

Insight mixed deep neural network architectures for molecular representation

Learning molecular representation is a crucial task in the field of drug discovery, particularly for various specific applications such as predicting molecular properties. Current methods are mainly based on deep neural network models, such as convolutional neural networks (CNN), recurrent neural networks (RNN), graph neural networks (GNN) and their mixed models. However, these neural network models mentioned above do not provide detailed explanations for the ability to learn molecular representations and why such experimental results occurred. In this paper, we aim to compare the performance of these models in predicting molecular properties based on molecular representation, and give more insight into deep neural network architectures for specific molecular tasks. Our experimental results demonstrate that the graph neural network can obtain superior performance on the regression tasks, while the mixed deep neural network models show better performance on the classification tasks. Ablation study also gives more explanation and analysis to the experimental results.

Image quality safety model for the safety of the intended functionality in highly automated agricultural machines

Achieving safe and reliable environmental perception is crucial for the success of highly automated or even autonomous agricultural machinery. However, developing such a system is challenging due to the inherent limitations of perception sensors. In certain conditions, these sensors may fail to capture accurate data, leading to erroneous perceptions of the environment and potentially compromising safety. Monitoring the functional insufficiencies of the measurement data is crucial for ensuring the safety and reliability of perception systems.This article introduces ISO standards, which provide guidelines for ensuring functional safety in highly automated mobile machines and vehicles. It also proposes an Image Quality Safety Model (IQSM) for monitoring the safety of the intended functionality in perception systems. The IQSM estimates the confidence level with which a camera can safely perform a specific object detection task. If the confidence level falls below a predefined threshold, the IQSM can trigger actions, alert operators, and prevent potential safety hazards. The IQSM exhibits remarkable performance, achieving a validation accuracy of about 90%, demonstrating its ability to effectively distinguish the safety of the intended functionality under a variety of image quality conditions.

MVP-HOT: A moderate visual prompt for hyperspectral object tracking

The growing attention to hyperspectral object tracking (HOT) can be attributed to the extended spectral information available in hyperspectral images (HSIs), especially in complex scenarios. This potential makes it a promising alternative to traditional RGB-based tracking methods. However, the scarcity of large hyperspectral datasets poses a challenge for training robust hyperspectral trackers using deep learning methods. Prompt learning, a new paradigm emerging in large language models, involves adapting or fine-tuning a pre-trained model for a specific downstream task by providing task-specific inputs. Inspired by the recent success of prompt learning in language and visual tasks, we propose a novel and efficient prompt learning method for HOT tasks, termed Moderate Visual Prompt for HOT (MVP-HOT). Specifically, MVP-HOT freezes the parameters of the pre-trained model and employs HSIs as visual prompts to leverage the knowledge of the underlying RGB model. Additionally, we develop a moderate and effective strategy to incrementally adapt the HSI prompt information. Our proposed method uses only a few (1.7M) learnable parameters and demonstrates its effectiveness through extensive experiments, MVP-HOT can achieve state-of-the-art performance on three hyperspectral datasets.

Measurement Guidance in Diffusion Models: Insight from Medical Image Synthesis.

In the field of healthcare, the acquisition of sample is usually restricted by multiple considerations, including cost, labor- intensive annotation, privacy concerns, and radiation hazards, therefore, synthesizing images-of-interest is an important tool to data augmentation. Diffusion models have recently attained state-of-the-art results in various synthesis tasks, and embedding energy functions has been proved that can effectively guide the pre-trained model to synthesize target samples. However, we notice that current method development and validation are still limited to improving indicators, such as Fréchet Inception Distance score (FID) and Inception Score (IS), and have not provided deeper investigations on downstream tasks, like disease grading and diagnosis. Moreover, existing classifier guidance which can be regarded as a special case of energy function can only has a singular effect on altering the distribution of the synthetic dataset. This may contribute to in-distribution synthetic sample that has limited help to downstream model optimization. All these limitations remind that we still have a long way to go to achieve controllable generation. In this work, we first conducted an analysis on previous guidance as well as its contributions on further applications from the perspective of data distribution. To synthesize samples which can help downstream applications, we then introduce uncertainty guidance in each sampling step and design an uncertainty-guided diffusion models. Extensive experiments on four medical datasets, with ten classic networks trained on the augmented sample sets provided a comprehensive evaluation on the practical contributions of our methodology. Furthermore, we provide a theoretical guarantee for general gradient guidance in diffusion models, which would benefit future research on investigating other forms of measurement guidance for specific generative tasks.

Hyperparameter recommendation via automated meta-feature selection embedded with kernel group Lasso learning

Hyperparameter recommendation via meta-learning relies on the characterization and quality of meta-features. These meta-features provide critical information about the underlying datasets but are often selected manually based on the practitioner’s experience and preference, which can be inefficient and ineffective in many applications. In this paper, we propose a novel hyperparameter recommendation approach that integrates with a Lasso-based multivariate kernel group (KGLasso) model. The developed KGLasso model automatically identifies primary meta-features through model training. By selecting the most explanatory meta-features for a specific meta-learning task, the recommendation performance becomes much more effective. Our KGLasso model builds on a group-wise generalized multivariate Lasso approach. Within this framework, we establish a minimization algorithm using a corresponding auxiliary function, which is mathematically proven to be convergent and robust. As an application, we develop a hyperparameter recommendation system using our built KGLasso model on 120 UCI datasets for the well-known support vector machine (SVM) algorithm. This system efficiently provides competent hyperparameter recommendations for new tasks. Extensive experiments, including comparisons with popular meta-learning baselines and search algorithms, demonstrate the superiority of our proposed approach. Our results highlight the benefits of integrating model learning and feature selection to construct an automated meta-learner for hyperparameter recommendation in meta-learning.

An effective surrogate-assisted rank method for evolutionary neural architecture search

Evolutionary neural architecture search (ENAS) is able to automatically design high-performed architectures of deep neural networks (DNNs) for specific tasks. In recent years, surrogate models have gained significant traction because they can estimate the performance of neural architectures, avoiding excessive computational costs for training. However, most existing surrogate models primarily predict the performance of architectures directly or predict pairwise comparison relationships, which makes it challenging to obtain the rank of a group of architectures when training samples are limited. To address this problem, we propose TCMR-ENAS, an effective triple-competition model-assisted rank method for ENAS. TCMR-ENAS employs a novel triple-competition surrogate model combined with a score-based fitness evaluation method to predict group performance rank. Moreover, a progressive online learning method is proposed to enhance the predictive performance of the triple-competition surrogate model in the framework of modified genetic search. To validate the effectiveness of TCMR-ENAS, we conducted a series of experiments on NAS-Bench-101, NAS-Bench-201, NATS-Bench and NAS-Bench-301, respectively. Experimental results show that TCMR-ENAS can achieve better performance with lower computational resources. The accuracies of searched architectures achieve the best results compared with those of the state-of-the-art methods with limited training samples. In addition, the factors that may influence the effectiveness of TCMR-ENAS are explored in the ablation studies.

Evolving Neural Network Designs with Genetic Algorithms: Applications in Image Classification, NLP, and Reinforcement Learning

A method of evolving deep learning architectures using genetic algorithms is presented. The method is a first step towards a low-cost evolutionary search for task-specific neural networks. We evolve task-specific model architectures optimized for fast execution and low error on several standard machine learning tasks: image classification, character-level language modeling, and solving the cart pole problem. We also introduce a simple variation of the method that is capable of evolving neural networks with recurrent connections of varying depth and length and show performance on a word-level language modeling task. The method is implemented in an open-source library. We hope that the ability to run an evolutionary search at this scale will make it possible for a wide audience to develop deep learning architectures that are specialized for a variety of tasks and to develop many interesting novel architectural features. A new method that uses evolutionary search to directly modify existing neural network architectures to perform a specific task is presented. We demonstrate that task-specific specialization of deep learning models can be useful in practice. We modify convolutional neural networks, residual networks, and an LSTM variant to perform various tasks, and show that specialized networks often perform better than models trained from scratch that have many more parameters and much larger training time. For example, on the object recognition task, a specialized model is built by training a base network to predict object position and then applying a series of genetic search operations to squeeze the network and fit new final layer weights to the output. The specialized model is 8 times faster and has 13% lower error, despite being 17 times smaller than a fully trained larger and slower network.

Multi-agent reinforcement learning with synchronized and decomposed reward automaton synthesized from reactive temporal logic

Multi-agent systems (MAS) consist of multiple autonomous agents interacting to achieve collective objectives. Multi-agent reinforcement learning (MARL) enhances these systems by enabling agents to learn optimal behaviors through interaction, thus improving their coordination in dynamic environments. However, MARL faces significant challenges in adapting to complex dependencies on past states and actions, which are not adequately represented by the current state alone in reactive systems. This paper addresses these challenges by considering MAS operating under task specifications formulated as Generalized Reactivity of rank 1 (GR(1)). These synthesized strategies are used as a priori knowledge to guide the learning. To tackle the difficulties of handling non-Markovian tasks in reactive systems, we propose a novel synchronized decentralized training paradigm that guides agents to learn within the MARL framework using a reward structure constructed from decomposed synthesized strategies of GR(1). We initially formalize the synthesis of GR(1) strategies as a reachability problem of winning states of the system. Subsequently, we develop a decomposition mechanism that constructs individual reward structures for decentralized MARL, incorporating potential values calculated through value iteration. Theoretical proofs are provided to verify that the safety and liveness properties are preserved. We evaluate our approach against other state-of-the-art methods under various GR(1) specifications and scenario maps, demonstrating superior learning efficacy and optimal rewards per episode. Additionally, we show that the decentralized training paradigm outperforms the centralized training paradigm. The value iteration strategy used to calculate potential values for the reward structure is compared against two other strategies, showcasing its advantages.

Leveraging Artificial Intelligence for Predictive Healthcare: A Data-Driven Approach to Early Diagnosis and Personalized Treatment

As noted, healthcare delivery systems are systems that always present organizational difficulties for any user and everyone related to these systems, but AI today is one of the widely-discussed and most frequently considered innovations that have prospects for radical improvement of the patient’s state and quality of life. AI solutions are on the verge of becoming integrated into clinical practice thanks to enhanced developments in the field, which means that there is a need to cover the part played by these technologies in healthcare and equip healthcare providers with the knowledge and resources needed to achieve that. This review article provides a broad, although dated, systematic synthesis of AI in clinical practice today, reviewing potential uses of AI in disease taxonomy and diagnosis, treatment suggestions, patient participation, and utilizing methodologies to consider the ethical and legal issues of AI and the persistent need for human insights and expertise. In order to establish the nature of the relationships and possible effects of AI in healthcare environments, the paper reviews the literature using PubMed/Medline, Scopus, and EMBASE databases. According to the findings, the application of AI technologies in health care delivery can greatly improve disease diagnosis, therapy choice, and clinical testing because, in contrast to human decision-making, AI algorithms analyze extensive data sets and outcompete human experts in specific tasks. It is accurate, cost-effective, time efficient, and free from human errors, advances the fields of pharmacogenomics and personalized medicine, optimizes the dosing of drugs, improves overall population health, establishes virtual healthcare services, and enhances mental health care, education, and trust between patients and physicians. Nevertheless, these and the following issues cannot be left unresolved: data privacy, AI bias, and the indispensability of the human factor.

Quantitative Evaluation for Robustness of Intelligent Fault Diagnosis Algorithms based on Self-attention Mechanism

Currently, various algorithmic models encounter numerous challenges in practical applications, such as noise, interference, and input changes, which can significantly impact their performance. Many methods have been proposed to enhance model robustness. However, to assess the effectiveness of these improvements, it is generally necessary to compare the model’s performance before and after applying the same noise and analyze the resulting changes. Moreover, evaluating the robustness of multiple models that meet basic requirements for a specific task requires qualitative analysis using specific indicators. This is especially crucial in fault diagnosis where multiple types of noise interference in the data can hinder accurate fault classification. Addressing this situation, this paper presents a quantitative evaluation method for the robustness of intelligent fault diagnosis algorithms based on the self-attention mechanism. The proposed method entails dividing the dataset into sub-datasets according to signal-to-noise ratio after injecting noise, separately calculating sub-indicators after training, dynamically assigning weights to these indicators using the self-attention mechanism and combining the weights of different sub-indicators to generate a comprehensive evaluation value for assessing robustness. The proposed method was validated through experiments involving three models, and the results demonstrated the reliability of this quantitative calculation approach for robustness.

Capital budgeting for electrical engineering projects: A practical methodology

The paper presents the capital budgeting process and identifies key elements in the process. A standardized approach in capital budgeting by application of information technology resources for both proposal submission and documentation becomes key to success in handling large volumes. An electrical engineering distribution system design project involves the application of engineering codes and standards to stay compliant and provide safe design and construction. A model starting with proposals acceptance to evaluation and approval of capital projects was presented. The proposal stage was decoded further into assessments of the existing electrical infrastructure based on aging and engineering standards. Upon review of the standards, the proposals were accepted for evaluation and approval. For business professionals, there is a gap when working in an electrical engineering construction industry when it comes to knowing applicable engineering standards and their application. This research filled that gap by enlisting the intent of the standards review process. Although the focus was on understanding a holistic approach to electrical engineering projects, a detailed analysis of specific tasks may be designed based on the model presented in this paper. Additionally, the transmission side of projects will have a similar application of capital budgeting.

NAS-CRE: Neural Architecture Search for Context-Based Relation Extraction

Relation extraction, a crucial task in natural language processing (NLP) for constructing knowledge graphs, entails extracting relational semantics between pairs of entities within a sentence. Given the intricacy of language, a single sentence often encompasses multiple entities that mutually influence one another. Recently, various iterations of recurrent neural networks (RNNs) have been introduced into relation extraction tasks, where the efficacy of neural network structures directly influences task performance. However, many neural networks necessitate manual determination of optimal parameters and network architectures, resulting in limited generalization capabilities for specific tasks. In this paper, we formally define the context-based relation extraction problem and propose a solution utilizing neural architecture search (NAS) to optimize RNN. Specifically, NAS employs an RNN controller to delineate an RNN cell, yielding an optimal structure to represent all relationships, thereby aiding in extracting relationships between target entities. Additionally, to enhance relation extraction performance, we leverage the XLNet pretrained model to comprehensively capture the semantic features of the sentence. Extensive experiments conducted on a real-world dataset containing words with multiple relationships demonstrate that our proposed method significantly enhances micro-F1 scores compared to state-of-the-art baselines.

Distributed Decision‐Making of General Linear Systems in Multi‐Coalition Games and Its Application to USV Swarm Confrontation

ABSTRACTThis article explores a decision‐making problem with partial information in a multi‐coalition game that involves both cooperation and competition. Considering heterogeneous players with general linear systems, we propose a distributed algorithm that employs average consensus mechanisms to estimate the gradient of the coalition function among players within each coalition and utilizes leader‐following protocols to estimate the actions of all players across multiple coalitions. Furthermore, to reduce communication costs, a dynamic event‐triggered mechanism (ETM) is introduced into the average consensus and leader‐following protocols. The ETM is asynchronous, eliminating the requirement for a global clock and allowing players to transmit information only when the triggering condition is satisfied. By the Lyapunov analysis, all actions asymptotically converge to the Nash equilibrium by the proposed algorithms. Moreover, the unmanned surface vehicle (USV) swarm confrontation is formulated as a coalition game. In the scenario of safeguarding territorial integrity, we devise specific tasks for both the invading and defending USVs, encompassing elements such as formation, intrusion, defense, and so forth. The effectiveness of the proposed algorithms is validated through comprehensive simulations.

Layer-Wise Learning Rate Optimization for Task-Dependent Fine-Tuning of Pre-Trained Models: An Evolutionary Approach

The superior performance of large-scale pre-trained models, such as Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformer (GPT), has received increasing attention in both academic and industrial research and has become one of the current research hotspots. A pre-trained model refers to a model trained on large-scale unlabeled data, whose purpose is to learn general language representation or features for fine-tuning or transfer learning in subsequent tasks. After pre-training is complete, a small amount of labeled data can be used to fine-tune the model for a specific task or domain. This two-stage method of “pre-training+fine-tuning” has achieved advanced results in natural language processing (NLP) tasks. Despite widespread adoption, existing fixed fine-tuning schemes that adapt well to one NLP task may perform inconsistently on other NLP tasks given that different tasks have different latent semantic structures. In this article, we explore the effectiveness of automatic fine-tuning pattern search for layer-wise learning rates from an evolutionary optimization perspective. Our goal is to use evolutionary algorithms to search for better task-dependent fine-tuning patterns for specific NLP tasks than typical fixed fine-tuning patterns. Experimental results on two real-world language benchmarks and three advanced pre-training language models show the effectiveness and generality of the proposed framework.

Domestic research on the history of chemistry

The presented review lists the main achievements and publications of Russian scientists in the field of the history of chemistry. The leading researches, major works, scientific centers are designated, the names of outstanding figures of science are named. The importance of the creation of the RAS Commission on the History of Chemistry in 2023 is emphasized, the main directions of its activities, specific tasks and the first results obtained are indicated.

Detection of Disease Features on Retinal OCT Scans Using RETFound

Eye diseases such as age-related macular degeneration (AMD) are major causes of irreversible vision loss. Early and accurate detection of these diseases is essential for effective management. Optical coherence tomography (OCT) imaging provides clinicians with in vivo, cross-sectional views of the retina, enabling the identification of key pathological features. However, manual interpretation of OCT scans is labor-intensive and prone to variability, often leading to diagnostic inconsistencies. To address this, we leveraged the RETFound model, a foundation model pretrained on 1.6 million unlabeled retinal OCT images, to automate the classification of key disease signatures on OCT. We finetuned RETFound and compared its performance with the widely used ResNet-50 model, using single-task and multitask modes. The dataset included 1770 labeled B-scans with various disease features, including subretinal fluid (SRF), intraretinal fluid (IRF), drusen, and pigment epithelial detachment (PED). The performance was evaluated using accuracy and AUC-ROC values, which ranged across models from 0.75 to 0.77 and 0.75 to 0.80, respectively. RETFound models display comparable specificity and sensitivity to ResNet-50 models overall, making it also a promising tool for retinal disease diagnosis. These findings suggest that RETFound may offer improved diagnostic accuracy and interpretability for specific tasks, potentially aiding clinicians in more efficient and reliable OCT image analysis.

Use of Drones for Trough Reading, Animal Counting, and Production Monitoring in Feedlot Systems

In line with the concept of precision agriculture, this study aimed to validate the use of digital aerial images captured using a remotely piloted aircraft (RPA) for collecting zootechnical data on cattle feedlot systems in a tropical environment. Images were captured on 21 non-consecutive days in 110 pens with up to 150 animals each. Conventional and RPA-based methods were adopted to determine animal behavior, feed trough levels, animal counts, and pen conditions. Data analysis revealed almost perfect agreement (kappa coefficient = 0.901) between trough readings taken by conventional and RPA methods as well as substantial agreement for fecal score (kappa coefficient = 0.785) and surface conditions (kappa coefficient = 0.737). However, animal counts and water quality scores showed only fair agreement, suggesting challenges in using RPA for these specific tasks. The results indicated that RPA represents a viable alternative to conventional methods for monitoring zootechnical indices in feedlots, offering benefits in terms of accuracy, efficiency, and cost-effectiveness. The implementation of RPA-based methods holds potential for improving animal management, welfare, and yield in feedlot systems.