Small-scale Networks Research Articles

Small-scale cross-layer fusion network for classification of diabetic retinopathy

Deep learning-based automatic classification of diabetic retinopathy (DR) helps to enhance the accuracy and efficiency of auxiliary diagnosis. This paper presents an improved residual network model for classifying DR into five different severity levels. First, the convolution in the first layer of the residual network was replaced with three smaller convolutions to reduce the computational load of the network. Second, to address the issue of inaccurate classification due to minimal differences between different severity levels, a mixed attention mechanism was introduced to make the model focus more on the crucial features of the lesions. Finally, to better extract the morphological features of the lesions in DR images, cross-layer fusion convolutions were used instead of the conventional residual structure. To validate the effectiveness of the improved model, it was applied to the Kaggle Blindness Detection competition dataset APTOS2019. The experimental results demonstrated that the proposed model achieved a classification accuracy of 97.75% and a Kappa value of 0.971 7 for the five DR severity levels. Compared to some existing models, this approach shows significant advantages in classification accuracy and performance.

Semantic-aware frame-event fusion based pattern recognition via large vision–language models

Pattern recognition through the fusion of RGB frames and Event streams has emerged as a novel research area in recent years. Current methods typically employ backbone networks to individually extract the features of RGB frames and event streams, and subsequently fuse these features for pattern recognition. However, we posit that these methods may suffer from two key issues: (1). They attempt to directly learn a mapping from the input vision modality to the semantic labels. This approach often leads to sub-optimal results due to the disparity between the input and semantic labels; (2). They utilize small-scale backbone networks for the extraction of RGB and Event input features, thus these models fail to harness the recent performance advancements of large-scale visual-language models. In this study, we introduce a novel pattern recognition framework that consolidates the semantic labels, RGB frames, and event streams, leveraging pre-trained large-scale vision–language models. Specifically, given the input RGB frames, event streams, and all the predefined semantic labels, we employ a pre-trained large-scale vision model (CLIP vision encoder) to extract the RGB and event features. To handle the semantic labels, we initially convert them into language descriptions through prompt engineering and polish using ChatGPT, and then obtain the semantic features using the pre-trained large-scale language model (CLIP text encoder). Subsequently, we integrate the RGB/Event features and semantic features using multimodal Transformer networks. The resulting frame and event tokens are further amplified using self-attention layers. Concurrently, we propose to enhance the interactions between text tokens and RGB/Event tokens via cross-attention. Finally, we consolidate all three modalities using self-attention and feed-forward layers for recognition. Comprehensive experiments on the HARDVS and PokerEvent datasets fully substantiate the efficacy of our proposed SAFE model. The source code has been released at https://github.com/Event-AHU/SAFE_LargeVLM.

Single and multi-threaded power flow algorithm for integrated transmission-distribution-residential networks

The rapid development of renewable sources has significantly increased interdependencies between electricity networks of all voltage levels, leading to bidirectional flows between transmission and distribution networks, and requiring analysis of Integrated Transmission-Distribution (ITD) network. The interconnectivity is further amplified by additional coupling with electricity, gas, heat and hydrogen networks, on both national and regional levels. Moreover, installation of solar, wind and storage on customers’ premises has indicated that residential networks need to be included in the overall integrated model, called Integrated Transmission-Distribution-Residential (ITDR) network. The main goal of the paper is to develop a general model of the ITD/ITDR networks and to solve the power-flow problem on large-scale networks in an efficient way. The general model includes three-phase transmission and multi-phase distribution models, as well as accurate control strategies for traditional and electronically coupled electricity resources. The proposed model is solved via novel single-threaded power flow procedure, which incorporates new network elements’ models and the developed algorithm for integrated power flow calculations in different domains. This is further improved by developing a multi-threaded approach. Analyses on a small-scale ITD network have shown that single- and multi-threaded approaches are, respectively, (1.5–4) and (2–5) times faster compared to the state-of-the-art procedures. As network size increases, the efficiency of the proposed multi-threaded procedure becomes more pronounced compared to the single-threaded one (up to 2.39 times).

Microfluidics-A novel technique for high-quality sperm selection for greater ART outcomes.

Microfluidics represent a quality sperm selection technique. Human couples fail to conceive and this is so in a significant population of animals worldwide. Defects in male counterpart lead to failure of conception so are outcomes of assisted reproduction affected by quality of sperm. Microfluidics, deals with minute volumes (μL) of liquids run in small-scale microchannel networks in the form of laminar flow streamlines. Microfluidic sperm selection designs have been developed in chip formats, mimicking invivo situations. Here sperms are selected and analyzed based on motility and sperm behavioral properties. Compared to conventional sperm selection methods, this selection method enables to produce high-quality motile sperm cells possessing non-damaged or least damaged DNA, achieve greater success of insemination in bovines, and achieve enhanced pregnancy rates and live births in assisted reproduction-in vitro fertilization (IVF) and intracytoplasmic sperm injection (ICSI). Besides, the concentration of sperm available to oocyte can be controlled by regulating the flow rate in microfluidic chips. The challenges in this technology are commercialization of chips, development of fully functional species-specific microfluidic tools, limited number of studies available in literature, and need of thorough understanding in reproductive physiology of domestic animals. In conclusion, incorporation of microfluidic system in assisted reproduction for sperm selection may promise a great success in IVF and ICSI outcomes. Future prospectives are to make this technology more superior and need to modify chip designs which is cost effective and species specific and ready for commercialization. Comprehensive studies in animal species are needed to be carried out for wider application of microfluidic sperm selection in invitro procedures.

Abnormal cascading dynamics in transportation networks based on Gaussian distribution of load

In the transportation network, we observe that the distance people travel by means of transportation follows a certain distribution. Statistical analysis shows that people’s travel distance is mainly concentrated in a medium range by the same vehicle, and they choose fewer destinations that are extremely close or far away. However, in previous studies, the impact of distance on the distribution of load flow within the network has often been neglected, or at best, addressed with overly simplistic assumptions. Therefore, we quantify the load flow distribution based on the Gaussian distribution of distances between the nodes. On this basis, a new cascading failure model is proposed using the shortest path strategy to calculate the initial load of the edge. Through the simulation of three real traffic networks and two artificially constructed networks with similar structural characteristics of traffic networks, we found the following interesting anomalies: First, increasing the load-bearing capacity of edges within the network does not necessarily lead to enhanced robustness. Second, we observed that removing more edges does not necessarily lead to a decrease in network robustness; conversely, the network robustness can be higher when a moderate number of edges are removed compared to fewer edges. To better understand the two anomalous dynamics phenomena we observed, we ran simulations on a small-scale network extracted from a real traffic network. We found that, under certain circumstances, the premature failure of some edges may isolate certain regions from the network, which may be responsible for this paradox.

Local detouredness: A new phenomenon for modelling route choice and traffic assignment

This study introduces the novel concept of local detouredness, i.e. detours on subsections of a route, as a new phenomenon for understanding and modelling route choice. Traditionally, Stochastic User Equilibrium (SUE) traffic assignment models have been concerned with judging the attractiveness of a route by its total route cost. However, through empirical analysis we show that considering solely the global properties of a route is insufficient. We find that it is important to consider local detouredness both when determining realistic and tractable route choice sets and when determining route choice probabilities. For example, analysis of observed route choice data shows that route usage tends to decay with local detouredness, and that there is an apparent limit on the amount of local detouredness seen as acceptable. No existing models can account for this systematically and consistently, which is the motivation for the new route choice model proposed in this paper: the Bounded Choice Model with Local Detour Threshold (BCM-LDT). The BCM-LDT model incorporates the effect of local detouredness on route choice probability, and has an in-built mechanism that assigns zero probabilities to routes violating a bound on total route costs and/or a threshold on local detouredness. Thereby, the model consistently predicts which routes are used and unused. Moreover, the probability expression is closed-form and continuous. SUE conditions for the BCM-LDT are given, and solution existence is proven. Exploiting the special structure of the problem, a novel solution algorithm is proposed where flow averaging is integrated with a modified branch-and-bound method that iteratively column-generates all routes satisfying local and global bounds. Numerical experiments are conducted on small-scale and large-scale networks, establishing that equilibrated solutions can be found and demonstrating the influence of the BCM-LDT parameters on choice set size and flow allocation.

Analysis of the Spatial Distribution and Common Mode Error Correlation in a Small-Scale GNSS Network.

When analyzing GPS time series, common mode errors (CME) often obscure the actual crustal movement signals, leading to deviations in the velocity estimates of station coordinates. Therefore, mitigating the impact of CME on station positioning accuracy is crucial to ensuring the precision and reliability of GNSS time series. The current approach to separating CME mainly uses signal filtering methods to decompose the residuals of the observation network into multiple signals, from which the signals corresponding to CME are identified and separated. However, this method overlooks the spatial correlation of the stations. In this paper, we improved the Independent Component Analysis (ICA) method by introducing correlation coefficients as weighting factors, allowing for more accurate emphasis or attenuation of the contributions of the GNSS network's spatial distribution during the ICA process. The results show that the improved Weighted Independent Component Analysis (WICA) method can reduce the root mean square (RMS) of the coordinate time series by an average of 27.96%, 15.23%, and 28.33% in the E, N, and U components, respectively. Compared to the ICA method, considering the spatial distribution correlation of stations, the improved WICA method shows enhancements of 12.53%, 3.70%, and 8.97% in the E, N, and U directions, respectively. This demonstrates the effectiveness of the WICA method in separating CMEs and provides a new algorithmic approach for CME separation methods.

Dynamics of a susceptible-infected-recovered model on complex networks with interregional migration.

We present a susceptible-infected-recovered model based on a dynamic flow network that describes the epidemic process on complex metapopulation networks. This model views population regions as interconnected nodes and describes the evolution of each region using a system of differential equations. The next-generation matrix method is used to derive the global basic reproduction number for three cases: a general network with homogeneous infection rates in all regions, a fully connected network, and a star network with heterogeneous infection and recovery rates. For the homogeneous case, we show that this global basic reproduction number is independent of the migration rates between regions. However, the rate of convergence of each region to an equilibrium state exhibits a much larger variance in random (Erdős-Rényi) networks compared to small-scale (Barabási-Albert) networks. For the general heterogeneous case, we report interesting results, namely that the global basic reproduction number decays exponentially with respect to the smallest nonzero Laplacian eigenvalue (algebraic connectivity). Furthermore, we demonstrate both analytically and numerically that as the network's algebraic connectivity increases, either by increasing the average node degree of each region or the global migration rate, the global basic reproduction number decreases and converges to the ratio of the average local infection rate to the average local recovery rate, meaning that the lower bound of the global basic reproduction rate does not equal the mean of local basic reproduction rates.

Multi-paradigm modelling and control of microgrid systems for better power stability in the Rockaways

The Rockaways Peninsula faces issues related to congestion and power outages during times of peak usage. Additionally, it is susceptible to disruptions caused by disasters such as hurricanes and storms. In this paper, we propose a new methodology that employs multi-paradigm modelling and control for the design and implementation of interconnected microgrid systems in the Rockaways. Microgrids are small-scale power networks that incorporate renewable energy technologies for power generation and distribution to enhance the control of energy supply and demand. Multi-paradigm modelling is employed to describe microgrids’ dynamic behavior more accurately by integrating system dynamics, agent-based modelling, as well as discrete event and continuous time simulation. We use agent-based models to describe the behavior of separate microgrid elements and the microgrid as a whole. Discrete event/continuous time simulation is used to analyze real-time operation of electrical parameters, such as voltage, current and frequency. Thus, the design, analysis and performance of microgrids are improved. Also, control strategies are used for the purpose of enabling the microgrids to operate effectively by responding to changes in power supply and demand and minimizing the effects of disturbances. The findings of this study demonstrate the feasibility and resilience benefits of incorporating multi-paradigm modelling and control in the design and management of microgrid systems in the Rockaways, which can result in the development of more durable, efficient, and sustainable energy systems in the region.

Passenger Route and Departure Time Guidance Under Disruptions in Oversaturated Urban Rail Transit Networks

The urban rail transit (URT) system attracts many commuters with its punctuality and convenience. However, it is vulnerable to disruptions caused by factors such as extreme weather and temporary equipment failures, which greatly affect passengers’ journeys and diminish the system’s service quality. In this study, we propose targeted travel guidance for passengers at different space–time locations by devising passenger rescheduling strategies during disruptions. This guidance not only offers insights into route changes but also provides practical recommendations for delaying departure times when required. We present a novel three-feature four-group passenger classification principle, integrating temporal, spatial, and spatiotemporal features to classify passengers in disrupted URT networks. This approach results in the creation of four distinct solution spaces based on passenger groups. A mixed integer programming model is built based on individual level considering the first-in-first-out rule in oversaturated networks. In addition, we present a two-stage solution approach for handling the complex issues in large-scale networks. Experimental results from both small-scale artificial networks and the real-world Beijing URT network validate the efficacy of our proposed passenger rescheduling strategies in mitigating disruptions. Specifically, when compared to scenarios with no travel guidance during disruptions, our strategies achieve a substantial reduction in total passenger travel times by 29.7% and 50.9%, respectively, underscoring the effectiveness in managing unexpected disruptions.

Investigation of the energy performance of thermal energy sharing for photovoltaic-thermal systems in a thermal prosumer network

This study analyzes the potential for energy saving through the integration of a photovoltaic-thermal (PVT) system with a heat pump (HP) in a community with a small-scale thermal network. An empirical study was conducted in a community to analyze and compare the performance of a HP and small-scale thermal network with hub thermal energy storage system that utilizes the surplus thermal energy produced by a PVT system installed in a residential house from other buildings with that of an existing electric hot water boiler. The energy saving analysis of the proposed system revealed that the energy consumption could be reduced by more than 70 % compared to the existing system. The HP exhibited a COP ranging between 4.2 and 4.7. Throughout the experiment, the integrated system collected an average of 23.77 kWh of heat energy daily from the PVT units. The annual domestic hot water load coverage from PVT through the thermal network was found to be 50.4 %. These results indicate that PVT systems combined with thermal networks and HPs can effectively reduce energy consumption in individual buildings and enhance energy efficiency within communities.

Firefighting Vehicle Dispatch and Route Planning Model Based on Priority Passage Rights

<div>Efficient fire rescue operations in urban environments are critical for saving lives and reducing property damage. By utilizing connected vehicle systems (CVS) for firefighting vehicles planning, we can reduce the response time to fires while lowering the operational costs of fire stations. This research presents an innovative nonlinear mixed-integer programming model to enhance fire rescue operations in urban settings. The model focuses on expediting the movement of firefighting vehicles within intricate traffic networks, effectively tackling the complexities associated with collaborative dispatch decisions and optimal path planning for multiple response units. This method is validated using a small-scale traffic network, providing foundational insights into parameter impacts. A case study in Sioux Falls shows its superiority over traditional “nearest dispatch” methods, optimizing both cost and response time significantly. Sensitivity analyses involving clearance speed, clearance time, minimum rescue force, and fire loss parameters contribute to the enhancement of urban fire rescue operations and the refinement of practical decision support systems.</div>

Effect of prediction uncertainties on the performance of a white-box model predictive controller for district heating networks

The performance of a model predictive controller (MPC) depends on the accuracy of the controller model and predictions of system disturbances. This paper studies the effect of prediction uncertainties and/or an imperfect state observer on the performance of a white-box MPC applied to a small-scale district heating network by performing a co-simulation with a detailed emulator model. This effect is studied for both a non-linear program (NLP)-based MPC and a mixed-integer NLP (MINLP)-based MPC. The results show a limited impact of prediction uncertainties when using a perfect state observer, with electricity use and thermal discomfort changing by only ▪ to ▪ and ▪/day/building to 0.56 ▪/day/building. An MPC with perfect predictions and a simple state observer only impacts the MPC's trade-off between electricity use and thermal discomfort: the former increases by 2–▪ but the latter decreases by 0.01–▪/day/building. The combined effect of both simplifications has a significant impact with an increase in electricity use and thermal discomfort of 2–▪ and 0.35–▪/day/building, but MPC still outperforms a conventional rule-based controller. The MINLP-based MPC better predicts the actual system behaviour, but its performance is similar compared to the NLP-based MPC in terms of electrical energy use and thermal comfort.

Physical layer encryption-based secure slicing in 5G RAN with hybrid-trusted links

In a 5G radio access network (RAN), network slicing enables dividing a single RAN infrastructure into multiple logical networks, efficiently accommodating services with diverse requirements. Although RAN slicing can help improve resource efficiency and reduce network costs, it is accompanied by various security risks. One of the security threats in RAN slicing is potential eavesdropping, resulting in the leakage of sensitive data within slices. Encryption technologies have been developed to address the eavesdropping problem at different layers in optical networks. We focus on physical layer encryption since it has been demonstrated beneficial in line-speed processing, low latency, and small encryption overhead. The problem of utilizing physical layer encryption technologies to achieve secure RAN slices remains unexplored since physical layer encryption introduces additional hardware costs. In this paper, we study how to realize secure RAN slicing based on physical layer encryption in a metro aggregation network that consists of hybrid-trusted links (i.e., links with different risks for eavesdropping). We propose an integer linear programming (ILP) model and an auxiliary graph-based heuristic for small-scale and large-scale networks, respectively. The objective is to maximize the number of deployed slices and minimize the total cost of secure slice deployment, which includes the costs of servers, line cards (LCs), encryption cards (ECs), and bandwidth resources. To evaluate the benefit of encryption, we compare it with a detour solution, which protects slices by routing through trusted links (i.e., where no additional hardware for encryption is deployed). Simulation results show that the encryption-based solution exhibits a lower cost than the benchmark when the same number of slices are deployed, and it can reduce the blocking ratio by up to 8.5% as slice requests increase. In addition, the average latency of slices is also reduced by up to 14.6%.

Shared autonomous vehicle operational decisions with vehicle movement and user travel behaviour

Shared Autonomous Vehicle (SAV) has many impacts on the transport development, such as saving parking space. However, SAV meets a huge challenge in terms of vehicle supply and user demand imbalance. The traditional mathematical optimization method cannot be well used due to the computational burden. Hence, this paper proposes a Reinforcement Learning (RL) based SAV relocation approach. First, two types of RL agents, car-based and zone-based agents, are developed as agents for vehicles and stations, respectively. Then, the RL scheme is trained by using historical demand data to facilitate real-time carsharing relocation. Finally, to compare the proposed two types of RL methods, three scenarios are used: small-scale, middle-scale, and large-scale networks. Solutions indicate that the enhanced zone-based method achieves an additional 146% profit compared to traditional threshold-based relocation strategies. The user travel behaviour impacts are provided by analyzing parking demand and travel movements among residential, industrial and commercial zones.

DTPBFT:A dynamic and highly trusted blockchain consensus algorithm for UAV swarm

With the continuous development of UAV technology, the application of UAV swarm in the military field is gradually becoming the focus of research around the world. Although it can bring a series of benefits in autonomous cooperation, the traditional UAV management technology is prone to hacker attacks due to many security issues such as a single point of failure brought by centralized management. Because of the advantages of distributed, tamper-proof, and traceability, blockchain is applied to UAV swarm to solve some of the security problems caused by centralized management. However, due to the limitations of its consensus algorithm Practical Byzantine Fault Tolerance (PBFT), its communication complexity will increase rapidly with the increase of the number of nodes, which also leads to the poor scalability of the algorithm and can only be applied to small-scale networks. To use the PBFT algorithm in large-scale networks such as UAV swarm, a dynamic and highly trusted PBFT (DTPBFT) algorithm is proposed in this paper. Firstly, a new consensus algorithm model including consensus layer and verification layer is designed. Then, a consensus node election scheme based on trust mechanism is proposed under this model. The trust degree of nodes in the blockchain network is comprehensively evaluated by selecting several representative indicators, and the weight factor of each indicator is calculated by entropy weight method. It not only reduces the communication complexity, but also ensures the reliability and dynamic update of consensus nodes. Experiments show that when the number of UAVs is 200, the consensus time of DTPBFT is 0.24 s, which indicates that this algorithm can support large-scale UAV swarm without causing communication congestion, so it has good scalability. In addition, experiments also show that DTPBFT can tolerate more than 13 malicious nodes, which improves the fault tolerance rate of PBFT.

Multiple rumor source identification in social networks leveraging community and monitor information

Multiple rumor source identification (MRSI) in social networks has become a challenging problem for controlling rumors from spreading automatically. Even though several techniques have been introduced for MRSI, most of them were introduced based on the fact that they knew the underlying diffusion model in advance, which is mostly not possible in real-world scenarios. So, this paper proposes a new algorithm called Multi-Source Detection, which uses Community and Monitor information (MSDCM). Our algorithm has two stages: detecting multiple rumor sources using the community and monitor information. First, identify the communities in the network to find the suspicious communities with the possible sources using the monitors’ information available in those communities. Next, detect a single source in each of the most likely suspicious communities by back-tracking from each monitor using edge weights in that community. Experimented on several datasets, including small-scale, large-scale, and artificial networks. Our results demonstrate that the designed algorithm is more efficient than the state-of-the-art algorithms.

Equator to Pole Solar Chromospheric Differential Rotation Using Ca-K Features Derived from Kodaikanal Data

Differential rotation is one of the basic characteristics of the Sun, and it plays an important role in generating the magnetic fields and its activities. We investigated rotation rate using chromospheric features such as plages, enhanced network (EN), active network (AN), and quiet network (QN) separately (for the first time). The digitized Ca-K images from Kodaikanal Observatory for 1907–1996 are used to study rotation over 0°–80° latitudes at an interval of 10°. We find that plages and all types of networks exhibit the differential rotation of the chromosphere. Furthermore, the rotation rate shows a decreasing pattern as one move from the equator to the higher polar latitudes for all the features used in the study. At the equator the rotation rate (rotation period) is obtained to be ∼13.98° day−1 (25.74 days), ∼13.91° day−1 (25.88 days), ∼13.99° day−1 (25.74 days), and ∼14.11° day−1 (25.51 days) for plage, EN, AN, and QN areas, respectively. By analyzing how the area of chromospheric features varies over time, we can effectively map the Sun’s rotation rate at all latitudes, including the polar regions. Interestingly, both plages and small-scale networks exhibit a similar differential rotation rate. This suggests these features likely rooted at the same layer below the visible surface of the Sun. Therefore, the long-term Ca-K data is very useful for studying the solar rotation rate at all latitudes including the polar regions.

Performance evaluation of six mesh router replacement methods for wireless mesh networks: A comparison study for small and middle scale networks considering two islands distribution of mesh clients

In this paper, we present a hybrid intelligent simulation system for optimization of mesh routers in Wireless Mesh Networks (WMNs) called WMN-PSOHCDGA. We implemented six mesh router replacement methods: CM, RIWM, LDIWM, LDVM, RDVM and FC-RDVM and consider Two Islands distribution of mesh clients. We carry out a comparison study of these router replacement methods for small and middle scale WMNs. We assessed the performance by computer simulations. The simulation results show that six methods have a good performance for connectivity and coverage metrics, for both small and middle scale WMNs. However, they have different behavior for load balancing. For small scale WMNs, the load balancing of LDIWM, RIWM and FC-RDVM is better than CM, LDVM and RDVM. While, comparing LDIWM, RIWM and FC-RDVM, the LDIWM has better load balancing. We found that the load balancing for small scale WMNs is not good, because there is a concentration of mesh routers in some areas. For middle scale WMNs, the CM, LDIWM, LDVM and RDVM have not a good load balancing. While, the RIWM and FC-RDVM have better performance. Comparing RIWM and FC-RDVM, we found that the load balancing of FC-RDVM is better than RIWM.

Algorithmic Foundation of Federated Learning with Sequential Data

The current analysis of federated optimization algorithms for training deep neural networks assumes that the data is non-sequential (e.g., images), which incurs a smooth loss objective. In contrast, edge devices generate lots of sequential data every day, where these sequences exhibit significant sequential correlation at different time stamps (e.g., text messages). In order to learn from such sequential data, people typically use a class of neural networks that is inherently nonsmooth, with a potentially unbounded smoothness parameter. Examples include recurrent neural networks, long-short-term memory networks, and transformers. It remains unclear how to design provably efficient algorithms for training these neural networks to learn from sequential data. My goal is to lay the algorithmic foundation of federated learning with sequential data, which contributes novel algorithms for learning from a range of real-world sequential data (e.g., natural language, electronic health record, transportation, time series, etc.) using state-of-the-art deep neural networks. In this talk, I will first motivate the problem by showing that the transformer, which is widely used for sequential data learning, has an unbounded smooth landscape. Then, I will introduce provably efficient federated deep learning algorithms in the presence of unbounded smoothness. In particular, I will introduce a few efficient algorithms for various settings of federated learning, including homogeneous data, heterogeneous data, and partial client participation. The main result is twofold. First, we show that the designed algorithms provably small computational and communication complexities. Second, we establish fundamental hardness results in the unbounded smoothness setting. Ultimately, I will discuss the future challenges of extending our research framework from small-scale neural networks to large language models.