Architecture Model Research Articles

A Neural Architecture Search CNN for Alzheimer’s Disease Classification

The evolution of automated machine learning (AutoML) is gradually reengineering the design of deep learning architectures for various imaging tasks. AutoML effectively develops model architectures and tunes hyperparameters through neural architecture search (NAS). Deep learning model architecture design is generally considered a tedious and time-consuming task that requires mastery skills to develop robust and better-performing models for imaging tasks. Again, the model's hyperparameters must be well-tuned to ensure optimal performances, which can be tedious and time-consuming if the hyperparameters are manually selected; using existing hyperparameter optimization algorithms can be expensive regarding resources. This study addresses these challenges in developing an optimal convolutional neural network (CNN) for classifying Alzheimer's (AD). The study, therefore, adopted a NAS approach to generate a CNN model architecture using a customized search space comprising only CNN patterns implemented with a NAS framework. The search was done for ten (10) trials, yielding a CNN architecture with an accuracy of 95.85% and a loss of 0.22. Training the model with a 10-fold cross-validation approach using a 0.0009 learning rate for 150 epochs improved the model's performance. The model recorded 97.17% accuracy, 97.21% precision, 97.14% recall, and a 0.99 area under the curve (AUC) in classifying AD as one of AD, mild cognitive impairment (MCI), and normal control (NC). The model obtained 98.06%, 98.66%, and 98.62% accuracy on binary classes of AD/NC, AD/MCI, and NC/MCI, respectively. The model generally showed robustness and better performance than existing CNN architectures.

Evolution and future directions of Artificial Intelligence Generated Content (AIGC): A comprehensive review

Artificial Intelligence Generated Content (AIGC) has rapidly evolved, revolutionizing the creation of text, images, audio, and video content. Despite these advancements, research on the development process of AIGC technology remains scarce, necessitating a systematic discussion of its current state and future directions. So this paper delves into the significant advancements and foundational technologies driving AIGC, emphasizing the contributions of state-of-the-art models such as DALL-E 3 [1] and Sora [2]. We analyze the evolution of generative models from single-modal approaches to the current multimodal generative models. The paper further explores the application prospects of AIGC across various domains such as office work, art, education, and film, while addressing the existing limitations and challenges in the field. We propose potential improvement directions, including more efficient model architectures and enhanced multimodal capabilities. Emphasis is placed on the environmental impact of AIGC technologies and the need for sustainable practices. Our comprehensive review aims to provide researchers and professionals with a deeper understanding of AIGC, inspiring further exploration and innovation in this transformative domain.

A community effort to optimize sequence-based deep learning models of gene regulation.

A systematic evaluation of how model architectures and training strategies impact genomics model performance is needed. To address this gap, we held a DREAM Challenge where competitors trained models on a dataset of millions of random promoter DNA sequences and corresponding expression levels, experimentally determined in yeast. For a robust evaluation of the models, we designed a comprehensive suite of benchmarks encompassing various sequence types. All top-performing models used neural networks but diverged in architectures and training strategies. To dissect how architectural and training choices impact performance, we developed the Prix Fixe framework to divide models into modular building blocks. We tested all possible combinations for the top three models, further improving their performance. The DREAM Challenge models not only achieved state-of-the-art results on our comprehensive yeast dataset but also consistently surpassed existing benchmarks on Drosophila and human genomic datasets, demonstrating the progress that can be driven by gold-standard genomics datasets.

State-of-the-Art Results with the Fashion-MNIST Dataset

In September 2024, the Fashion-MNIST dataset will be 7 years old. Proposed as a replacement for the well-known MNIST dataset, it continues to be used to evaluate machine learning model architectures. This paper describes new results achieved with the Fashion-MNIST dataset using classical machine learning models and a relatively simple convolutional network. We present the state-of-the-art results obtained using the CNN-3-128 convolutional network and data augmentation. The developed CNN-3-128 model containing three convolutional layers achieved an accuracy of 99.65% in the Fashion-MNIST test image set. In addition, this paper presents the results of computational experiments demonstrating the dependence between the number of adjustable parameters of the convolutional network and the maximum acceptable classification quality, which allows us to optimise the computational cost of model training.

Advances in medical image analysis: A comprehensive survey of lung infection detection

AbstractThis research investigates advanced approaches in medical image analysis, specifically focusing on segmentation and classification techniques, as well as their integration into multi‐task architectures for lung infections. This research begins by explaining key architectural models used in segmentation and classification tasks. The study extends to the enhancement of these architectures through attention modules and conditional random fields. Relevant datasets and evaluation metrics, incorporating discussions on loss functions are also reviewed. This review encompasses recent advancements in single‐task and multi‐task models, highlighting innovations in semi‐supervised, self‐supervised, few‐shot, and zero‐shot learning techniques. Empirical analysis is conducted on both single‐task and multi‐task architectures, predominantly utilizing the U‐Net framework, and is applied across multiple datasets for segmentation and classification tasks. Results demonstrate the effectiveness of these models and provide insights into the strengths and limitations of different approaches. This research contributes to improved detection and diagnosis of lung infections by offering a comprehensive overview of current methodologies and their practical applications.

Non-blocking functional dependency discovery from data streams

With the proliferation of sensors and IoT technologies, there is an increasing need to analyze information from data streams that they produce dynamically. However, the volume and velocity of this data require algorithms that mine knowledge as data are read from streams. The capability of dynamically extracting functional dependencies (fds) from data streams would not only permit to assess and improve the quality of data, but also provide knowledge on the evolution of data correlations within streams, allowing to understand the relevance that each feature has in predicting unknown features. In this paper, we propose a new discovery algorithm, namely COD3, which allows to continuous discovery fds holding on a data stream, as the data are read from it. COD3 represents the first proposal to use a non-blocking architectural model for discovering fds from data streams. Furthermore, we present novel data structures and a validation method to handle dynamic discovery and reduce data load inbound streams. Experimental evaluations demonstrate its effectiveness on both adapted real-world datasets and real data streams, such as those from air quality sensors. Moreover, by integrating COD3 with Bleach, a well-known fd-based data stream cleansing framework, we demonstrate its effectiveness in a real-world use case.

An inductive bias from quantum mechanics: learning order effects with non-commuting measurements

There are two major approaches to building good machine learning algorithms: feeding lots of data into large models or picking a model class with an “inductive bias” that suits the structure of the data. When taking the second approach as a starting point to design quantum algorithms for machine learning, it is important to understand how mathematical structures in quantum mechanics can lead to useful inductive biases in quantum models. In this work, we bring a collection of theoretical evidence from the quantum cognition literature to the field of quantum machine learning to investigate how non-commutativity of quantum observables can help to learn data with “order effects,” such as the changes in human answering patterns when swapping the order of questions in a survey. We design a multi-task learning setting in which a generative quantum model consisting of sequential learnable measurements can be adapted to a given task — or question order — by changing the order of observables, and we provide artificial datasets inspired by human psychology to carry out our investigation. Our first experimental simulations show that in some cases the quantum model learns more non-commutativity as the amount of order effect present in the data is increased and that the quantum model can learn to generate better samples for unseen question orders when trained on others — both signs that the model architecture suits the task.

A Brief Review of Lightweighting Methods for Vision Transformers (ViT)

The Vision Transformer (ViT) has emerged as a powerful model in recent years, surpassing traditional Convolutional Neural Networks (CNNs) in various benchmarks. However, its large model architecture and high parameter count present significant challenges for mobile deployment. Consequently, the exceptional performance of ViT in computer vision tasks is often overshadowed by its difficulties in being deployed on mobile devices due to its large parameter size and high computational demands. This paper provides a comprehensive review of the literature on lightweight ViT models, focusing on model optimization strategies such as post-training modifications—quantization, pruning, and knowledge distillation—as well as architectural changes including hybrid CNN-Transformer, MLP-based, and sparse models. These strategies are aimed at improving efficiency for mobile platforms. The review aims to clarify current techniques for mobile ViT, guide future research, stimulate innovation, and contribute to the development of efficient ViT models for mobile environments.

From Noise to Knowledge: Diffusion Probabilistic Model-Based Neural Inference of Gene Regulatory Networks.

Understanding gene regulatory networks (GRNs) is crucial for elucidating cellular mechanisms and advancing therapeutic interventions. Original methods for GRN inference from bulk expression data often struggled with the high dimensionality and inherent noise in the data. Here we introduce RegDiffusion, a new class of Denoising Diffusion Probabilistic Models focusing on the regulatory effects among feature variables. RegDiffusion introduces Gaussian noise to the input data following a diffusion schedule and uses a neural network with a parameterized adjacency matrix to predict the added noise. We show that using this process, GRNs can be learned effectively with a surprisingly simple model architecture. In our benchmark experiments, RegDiffusion shows superior performance compared to several baseline methods in multiple datasets. We also demonstrate that RegDiffusion can infer biologically meaningful regulatory networks from real-world single-cell data sets with over 15,000 genes in under 5 minutes. This work not only introduces a fresh perspective on GRN inference but also highlights the promising capacity of diffusion-based models in the area of single-cell analysis. The RegDiffusion software package and experiment data are available at https://github.com/TuftsBCB/RegDiffusion.

Tree-based classification model for Long-COVID infection prediction with age stratification using data from the national COVID cohort collaborative (N3C)

Abstract Objectives We propose and validate a domain knowledge-driven classification model for diagnosing Post-acute sequelae of SARS-CoV-2 infection (PASC), also known as Long COVID, using Electronic Health Records (EHR) data. Materials and Methods We developed a robust model that incorporates features strongly indicative of PASC or associated with the severity of COVID-19 symptoms as identified in our literature review. The XGBoost tree-based architecture was chosen for its ability to handle class-imbalanced data and its potential for high interpretability. Using the training data provided by the Long COVID Computation Challenge (L3C), which was a sample of the National COVID Collaborative (N3C), our models were fine-tuned and calibrated to optimize Area Under the Receiver Operating Characteristic curve (AUROC) and the F1 score, following best practices for the class-imbalanced N3C data. Results Our age-stratified classification model demonstrated strong performance with an average 5-fold cross-validated AUROC of 0.844 and F1 score of 0.539 across the young adult, mid-aged, and older-aged populations in the training data. In an independent testing dataset, which was made available after the challenge was over, we achieved an overall AUROC score of 0.814 and F1 score of 0.545. Discussion The results demonstrated the utility of knowledge-driven feature engineering in a sparse EHR data and demographic stratification in model development to diagnose a complex and heterogeneously presenting condition like PASC. The model's architecture, mirroring natural clinician decision-making processes, contributed to its robustness and interpretability, which are crucial for clinical translatability. Further, the model’s generalizability was evaluated over a new cross-sectional data as provided in the later stages of the L3C challenge. Conclusion The study proposed and validated the effectiveness of age-stratified, tree-based classification models to diagnose PASC. Our approach highlights the potential of machine learning in addressing the diagnostic challenges posed by the heterogeneity of Long COVID symptoms.

Deep Learning for Robust Facial Expression Recognition: A Resilient Defense Against Adversarial Attacks

Adversarial attacks can be extremely dangerous, particularly in scenarios where the precision of facial expression identification is of utmost importance. Hiring adversarial training methods proves effective in mitigating these threats. Although effective, this technique requires large computing resources. This study aims to strengthen deep learning model resilience against adversarial attacks while optimizing performance and resource efficiency. Our proposed method uses adversarial training techniques to create adversarial examples, which are permanently stored as a separate dataset. This strategy helps the model learn and enhances its resilience to adversarial attacks. This study also evaluates models by subjecting them to adversarial attacks, such as the One Pixel Attack and the Fast Gradient Sign Method, to identify any potential vulnerabilities. Moreover, we use two different model architectures to see how well they are protected against adversarial attacks. It compared their performances to determine the best model for making systems more resistant while still maintaining good performance. The findings show that the combination of the proposed adversarial training technique and an efficient model architecture outcome in increased resistance to adversarial attacks. This also improves the reliability of the model and saves more resources for computation. This is evidenced by the high accuracy results achieved at 98.81% accuracy on the CK+ datasets. The adversarial training technique proposed in this study offers an efficient alternative to overcome the limitations of computational resources. This fortifies the model against adversarial attacks, resulting in significant increases in model resilience without loss of performance.

Using total-body PET to investigate relationships between excretory organs time activity curves and the metabolite fraction for [18F]SynVesT-1

Accurate quantification of dynamic PET data using kinetic modelling requires assessment of the metabolism of the tracer to obtain the quantity of unchanged parent tracer in the blood. Several image derived input functions methods have been demonstrated to be effective for kinetic modelling; however, currently no image derived metabolite correction method exists. Finding an image derived metabolite correction would remove the need for invasive blood sampling. This study uses 7 full body 18F-SynVesT-1 total body mouse scans. Initially, the time activity curves (TACs) for a range of organs including the liver, kidney and bladder were studied. Factor analysis (FA) was investigated to accurately separate contributions from the kidney – cortex, medulla and collecting duct. Following this, the relationships between each of the organ TACs and the parent plasma fraction (PPf) was investigated. FA analysis successfully discriminates between contributions from the cortex, medulla and collecting duct of the kidneys. In the context of total body scans, FA removes the need for manual drawing of volumes of interest allowing for more consistency amongst studies. By separating out these regions, a relationship between the kidney medulla and the metabolites also came to light. In addition, the average of the seven bladder time activity curves against the plasma parent fraction [Figure 1] shows a potential relationship between the excretion of the tracer and its metabolism. Combining excretory organ relationships with the parent fraction and potentially liver data opens the possibility for a model to be produced to fit a metabolite curve when we have access to a total body PET scan. While this research is in its early stages, it has shown relationships between excretory organ TACs and the metabolite fraction. This has the potential to allow for metabolite correction models to be created directly from the image data. Model architectures are under investigation. Please click on the 'PDF' for the full abstract!

FaceDefend: Copyright Protection to Prevent Face Embezzle

With the rapid evolution of deep learning and the advent of artificial intelligence, the metaverse has emerged as a significant technology. Within the metaverse, diverse elements such as rich applications and realistic digital avatars provide users with immersive experiences, but it poses a series of security problems. Current research predominantly focuses on the data storage and transmission processes from the perspective of blockchain and the Internet of Things to achieve the protection of the metaverse. However, there exists a gap in security research on the digital avatar generation process. Given that digital avatars are the primary entities engaging in social activities within the metaverse and are crafted based on real face images, the virtual character can be generated easily by stealing the user’s face image and controlled to interact with others. In order to deal with the above problems, we propose a novel method to prevent the misuse of faces, which maintains the security of the metaverse by protecting facial data and thus preventing its misuse. We explore the common architecture of generative models and propose a defense method based on copyright protection to prevent face embezzling. Firstly, we utilize the copyright protection module to obtain copyright protection information. Secondly, we utilized the defense control module to ensure the representation of the protected images occurs errors in the latent space of the generation model. Therefore the subsequent generation task output fails, which effectively protects the face data and prevents the generation of digital avatars. Furthermore, the results on public datasets and across multiple generative models present unnatural outputs, indicating the excellence of our defense and transfer capabilities.

Experiencia didáctica para el estudio y valoración de la Arquitectura Vernácula guatemalteca

The didactic experience for the study and valoration of Guatemalan Vernacular Architecture consists of a series of academic activities in progressive phases, which introduce this type of architectura in danger of extinction to 4th grade students of architecture at the Rafael Landivar University, in the course: valoration of architecture and territory. Students are progressively immersed in the topic, from the bibliography to field visits, endingin the scale representation of an architectural model. This experience allows participant to learn about architecture, its compositional elements, its construction technique and materials, all whitin the framework of the appreciation of material and intangible cultural heritage. ILIA: Investigaciones Latinoamericanas en Ingeniería y Arquitectura, No. 01, 2024: 75-78.

Real-time robust liver and gallbladder segmentation during laparoscopic cholecystectomy using convolutional neural networks: an analysis

Aim: Images in different laparoscopic cholecystectomy datasets are acquired using various camera models, parameters, and settings, with the annotation methods varying by institution. These factors result in inconsistent inference performance of the network model. This study aims to identify the optimal network model architecture for liver and gallbladder segmentation from several options. Then, the performance and robustness of the optimal network model are evaluated using an independent dataset that is not included in the training. Methods: The public dataset, CholecSeg8k, was utilized as the input for the network model training, validation, and testing. A local private dataset from KPJ Damansara Hospital, Selangor, Malaysia, was used for testing purposes only. For the implementation of liver and gallbladder segmentation, segmentation models, a public Python library was employed. Results: Among the experiments, highly accurate liver and gallbladder segmentation results were achieved using the feature pyramid network (FPN) architecture as the network model, with the Inception-ResNet-v2 architecture as the network backbone. The best-trained network model resulted in a loss of 0.070955, a mean intersection over union (IoU) score of 0.95896, and a mean F1-score of 0.9773 on the test set. However, visualized results for the private dataset contained considerable false-negative areas. Conclusion: The proposed automated technique has the potential to serve as an alternative to the conventional indocyanine green injection along with near-infrared fluorescence imaging (ICG-NIRF)-based method for liver and gallbladder segmentation during laparoscopic cholecystectomy. Future work will focus on enhancing the results of the private dataset. Additionally, a surgeon-assistant robotic arm that will use the liver and gallbladder segmentation results for camera steering will be analyzed.

Identifying Re-identification Challenges: Past, Current and Future Trends

Person and vehicle re-identification has been a popular subject in the field of the computer vision technologies. Existing closed-set re-identification surpasses human-level accuracies on commonly used benchmarks, and the research focus for re-identification is shifting to the open world-setting. The latter setting is more suitable for practical applications, however, is less developed due to its challenges. On the other hand, existing research is more focused on person re-identification, even though both, person and vehicle, are important components for smart city applications. This review attempts to combine for the first time the problem of person and vehicle re-identification under closed and open settings, its challenges, and the existing research. Specifically, we start from the origin of the re-identification task and then summarize state-of-the-art research based on deep learning in different scenarios: person or vehicle or unified re-identification in closed- and open-world settings. Additionally, we analyse a new method for solving the re-identification task using the Transformer, a model architecture that relies entirely on an attention mechanism, which shows promising results. This survey facilitates future research by providing a summary on past and present trends, and aids to improve the usability of re-ID techniques.

LightGBM hybrid model based DEM correction for forested areas.

The accuracy of digital elevation models (DEMs) in forested areas plays a crucial role in canopy height monitoring and ecological sensitivity analysis. Despite extensive research on DEMs in recent years, significant errors still exist in forested areas due to factors such as canopy occlusion, terrain complexity, and limited penetration, posing challenges for subsequent analyses based on DEMs. Therefore, a CNN-LightGBM hybrid model is proposed in this paper, with four different types of forests (tropical rainforest, coniferous forest, mixed coniferous and broad-leaved forest, and broad-leaved forest) selected as study sites to validate the performance of the hybrid model in correcting COP30DEM in different forest area DEMs. In the hybrid model of this paper, the choice was made to use the Densenet architecture of CNN models with LightGBM as the primary model. This choice is based on LightGBM's leaf-growth strategy and histogram linking methods, which are effective in reducing the data's memory footprint and utilising more of the data without sacrificing speed. The study uses elevation values from ICESat-2 as ground truth, covering several parameters including COP30DEM, canopy height, forest coverage, slope, terrain roughness and relief amplitude. To validate the superiority of the CNN-LightGBM hybrid model in DEMs correction compared to other models, a test of LightGBM model, CNN-SVR model, and SVR model is conducted within the same sample space. To prevent issues such as overfitting or underfitting during model training, although common meta-heuristic optimisation algorithms can alleviate these problems to a certain extent, they still have some shortcomings. To overcome these shortcomings, this paper cites an improved SSA search algorithm that incorporates the ingestion strategy of the FA algorithm to increase the diversity of solutions and global search capability, the Firefly Algorithm-based Sparrow Search Optimization Algorithm (FA-SSA algorithm) is introduced. By comparing multiple models and validating the data with an airborne LiDAR reference dataset, the results show that the R2 (R-Square) of the CNN-LightGBM model improves by more than 0.05 compared to the other models, and performs better in the experiments. The FA-SSA-CNN-LightGBM model has the highest accuracy, with an RMSE of 1.09 meters, and a reduction of more than 30% of the RMSE when compared to the LightGBM and other hybrid models. Compared to other forested area DEMs (such as FABDEM and GEDI), its accuracy is improved by more than 50%, and the performance is significantly better than other commonly used DEMs in forested areas, indicating the feasibility of this method in correcting elevation errors in forested area DEMs and its significant importance in advancing global topographic mapping.

Real-time Blood Pressure Prediction on Wearable Devices using Edge Based Deep Neural Networks: A Hardware-software Co-design Approach

This paper presents the hardware realization of a real-time blood pressure (BP) prediction model for wearable devices, utilizing long short-term memory (LSTM) deep neural networks (DNNs). The proposed system uses both electrocardiogram (ECG) and photoplethysmogram (PPG) signal values for BP prediction. It aims to address the limitations of traditional BP measurement methods, providing a low error, minimal computational overhead, more accurate and convenient alternative system for individuals with hypertension or at risk for cardiovascular diseases. The utilization of split matrix approach leads to a reduction in hardware complexity across the entire system. This technique involves breaking down the larger weight matrices used in the computations of DNNs into smaller matrices. This fragmentation results in a decrease in the complexity of the hardware responsible for performing matrix vector multiplications (MVMs) within LSTMs. The resultant architecture of the predictive model gains several advantages, including a lowered level of complexity in terms of the space occupied by individual cells, decreased processing delay, and reduced power consumption. Furthermore, this approach enables the achievement of a notably improved minimum achievable clock period of 2.972 ns. This prediction model can operate locally on wearable devices, reducing the reliance on cloud computing and improving privacy and security. The performance evaluations are carried out using both analytical and implementation results. The results indicate that the proposed model can be practically applied to real-world problems and can potentially enhance the accuracy of various machine-learning tasks.

Assessing groundwater level modelling using a 1-D convolutional neural network (CNN): linking model performances to geospatial and time series features

Abstract. Groundwater level (GWL) forecasting with machine learning has been widely studied due to its generally accurate results and low input data requirements. Furthermore, machine learning models for this purpose can be set up and trained quickly compared to the effort required for process-based numerical models. Despite demonstrating high performance at specific locations, applying the same model architecture to multiple sites across a regional area can lead to varying accuracies. The reasons behind this discrepancy in model performance have been scarcely examined in previous studies. Here, we explore the relationship between model performance and the geospatial and time series features of the sites. Using precipitation (P) and temperature (T) as predictors, we model monthly groundwater levels at approximately 500 observation wells in Lower Saxony, Germany, applying a 1-D convolutional neural network (CNN) with a fixed architecture and hyperparameters tuned for each time series individually. The GWL observations range from 21 to 71 years, resulting in variable test and training dataset time ranges. The performances are evaluated against selected geospatial characteristics (e.g. land cover, distance to waterworks, and leaf area index) and time series features (e.g. autocorrelation, flat spots, and number of peaks) using Pearson correlation coefficients. Results indicate that model performance is negatively influenced at sites near waterworks and densely vegetated areas. Longer subsequences of GWL measurements above or below the mean negatively impact the model accuracy. Besides, GWL time series containing more irregular patterns and with a higher number of peaks might lead to higher model performances, possibly due to a closer link with precipitation dynamics. As deep learning models are known to be black-box models missing the understanding of physical processes, our work provides new insights into how geospatial and time series features link to the input–output relationship of a GWL forecasting model.

A bidirectional pedestrian macroscopic speed model construction based on pedestrian microscopic simulation experiments.

Studies of macroscopic speed modeling of bidirectional pedestrian cross-flows have relied heavily on scenario experiments, but the data itself may be deficient because large-scale scenario experiments are not easy to organize and subjects may not be walking under normal conditions. In order to explore the possibility of using microscopic pedestrian flow simulations for macroscopic speed modeling of pedestrian flows, a series of two-way pedestrian cross-flow simulation experiments were designed. Bidirectional pedestrian flows are defined as Peds1 and Peds2. The crossing angle and pedestrian flow rate are used as variables, and a bidirectional pedestrian flows simulation is designed as an orthogonal experiment. The crossing angles range from 15 to 165 degrees, and bidirectional pedestrian flow rate range from 1 ped/s to 8 ped/s. A series of simulations are built and performed on the GIS agent-based modeling architecture (GAMA) platform. By analyzing the flow data of bidirectional flows in the crossing area, it is found that when the Peds1 density falls below a threshold, Peds1 speed is determined by pedestrians themselves and mainly remains in a free flow state; otherwise, the Peds1 speed decreases with density. The clear effects such as Peds2 density on the Peds1 speed cannot be determined. A piecewise function combined with a linear function and an exponential function is constructed as the Peds1 speed model considering the influence of the crossing angle. The calibration results show that the piecewise function should be better than the non-piecewise function. Compared to the results of established studies, the results in this paper have some differences. Therefore, the simulation method cannot completely replace the scene experiments. However, this approach can provide suggestions for subsequent refinement of the experimental program, as well as a feasible direction for the construction of a speed relationship for bidirectional pedestrian flows.