Number of necessary training examples for Neural Networks with different number of trainable parameters

Götz, Theresa; Göb, St.; Sawant, Shrutika S.; Erick, X.F.; Wittenberg, Thomas; Schmidkonz, Christian; Tomé, Ana; Lang, Elmar; Ramming, Andreas

doi:10.1016/j.jpi.2022.100114

Cited by 9 publications

(4 citation statements)

References 54 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…1, a systematic approach was undertaken within a dedicated repository of image datasets of disasters and emergencies [38]. Next, global variables crucial for model training were modified, encompassing parameters such as the generic seed, number of epochs, learning rates, choice of pre-trained base models including EfficientNetB0, B7, V2B0, and V2L, InceptionV3, ResNet50, and VGG19, together with the pre-processing methods and optimization algorithms like Adam and RMSprop [39]. Subsequently, the dataset was read and decoded into pairs, ensuring proper preparation for training [40].…”

Section: Methodsmentioning

confidence: 99%

Exploring the Landscape: Analysis of Model Results on Various Convolutional Neural Network Architectures for iRESPOND System

Prianes,

Fortuno,

Onesa

et al. 2024

IJACSA

View full text Add to dashboard Cite

In the era of rapid technological advancement, the integration of cutting-edge technologies plays a pivotal role in enhancing the efficiency and responsiveness of critical systems. iRESPOND, a real-time Geospatial Information and Alert System, stands at the forefront of such innovations, facilitating timely and informed decision-making in dynamic environments. As the demand for accurate and swift responses, the role of CNN models in iRESPOND becomes significant. The study focuses on seven prominent CNN architectures, namely EfficientNet (B0, B7, V2B0, and V2L), InceptionV3, ResNet50, and VGG19 and with the integration of different optimizers and learning rates. The methodology employed a strategic implementation of looping during the training phase. This iterative approach is designed to systematically re-train the CNN models, emphasizing identifying the most suitable architecture among the seven considered variants. The primary objective is to discern the optimal architecture and fine-tune critical parameters, explicitly targeting the optimizer and learning rate values. The differential impact of each model on the system's ability is to discern patterns and anomalies in the image datasets. ResNet50 exhibited robust performance showcasing suitability for real-time processing in dynamic environments with a better accuracy result of 95.02%. However, the EfficientNetV2B0 model, characterized by its advancements in network scaling, presented promising results with a lower loss of 0.187. Generally, the findings not only contribute valuable insights into the optimal selection of architectures for iRESPOND but also highlight the importance of fine-tuning hyperparameters through an iterative training approach, which paves the way for the continued enhancement of iRESPOND as an adaptive system.

show abstract

Section: Methodsmentioning

confidence: 99%

Exploring the Landscape: Analysis of Model Results on Various Convolutional Neural Network Architectures for iRESPOND System

Prianes,

Fortuno,

Onesa

et al. 2024

IJACSA

View full text Add to dashboard Cite

show abstract

“…During the training process, the CNN was trained using 70% of the available images for training purposes and the remaining 30% for validation, which is considered by the literature as the optimal partition of the dataset for training these types of CNNs [35,76].…”

Section: Training Model Process and Testingmentioning

confidence: 99%

Remote Measurement of Tide and Surge Using a Deep Learning System with Surveillance Camera Images

Sabato,

Scardino,

Kushabaha

et al. 2024

Water

View full text Add to dashboard Cite

The latest progress in deep learning approaches has garnered significant attention across a variety of research fields. These techniques have revolutionized the way marine parameters are measured, enabling automated and remote data collection. This work centers on employing a deep learning model for the automated evaluation of tide and surge, aiming to deliver accurate results through the analysis of surveillance camera images. A mode of deep learning based on the Inception v3 structure was applied to predict tide and storm surges from surveillance cameras located in two different coastal areas of Italy. This approach is particularly advantageous in situations where traditional tide sensors are inaccessible or distant from the measurement point, especially during extreme events that require accurate surge measurements. The conducted experiments illustrate that the algorithm efficiently measures tide and surge remotely, achieving an accuracy surpassing 90% and maintaining a loss value below 1, evaluated through Categorical Cross-Entropy Loss functions. The findings highlight its potential to bridge the gap in data collection in challenging coastal environments, providing valuable insights for coastal management and hazard assessments. This research contributes to the emerging field of remote sensing and machine learning applications in environmental monitoring, paving the way for enhanced understanding and decision-making in coastal regions.

show abstract

“…In addition to data stream combinations, network architecture and training data set size also influence CNN performance. Networks with higher numbers of trainable parameters require more data during training to reduce the effects of overfitting (Götz et al., 2022). Network performance was investigated for PHD‐CNN architectures of varying complexity.…”

Section: Effect Of Network Architecture Complexity and Data Quantity ...mentioning

confidence: 99%

Parallel heterogeneous data‐fusion convolutional neural networks for improved rail bridge strike detection

Khresat,

Sitton,

Story

2024

Computer aided Civil Eng

View full text Add to dashboard Cite

Low clearance rail bridges provide vital crossings for freight and passenger trains and are susceptible to frequent strikes from overheight vehicles or equipment. Impact detection systems can help ensure the safety of railroad bridges and their users; such systems streamline monitoring efforts by providing near real‐time strike notifications to rail managers responsible for assessing a bridge after a strike. This paper develops parallel heterogeneous data‐fusion convolutional neural networks (PHD‐CNN) operating on data collected from in‐service rail bridges that improves detection and classification of from overheight vehicles. Convolutional neural networks (CNNs) automatically extract features from multiple data streams from different sensor modalities. The method provides a mechanism to homogenize and fuse disparate data streams for use as inputs to a classifier that distinguishes bridge strikes from passing trains. The study also provides practical implementation guidelines through a framework sensitivity characterization to examine the effects on performance of input data stream type, data set size, and CNN architecture complexity. Optimum networks detect, on average, 95% of bridge strikes with false positive rates less than 2%.

show abstract

Number of necessary training examples for Neural Networks with different number of trainable parameters

Cited by 9 publications

References 54 publications

Exploring the Landscape: Analysis of Model Results on Various Convolutional Neural Network Architectures for iRESPOND System

Exploring the Landscape: Analysis of Model Results on Various Convolutional Neural Network Architectures for iRESPOND System

Remote Measurement of Tide and Surge Using a Deep Learning System with Surveillance Camera Images

Parallel heterogeneous data‐fusion convolutional neural networks for improved rail bridge strike detection

Contact Info

Product

Resources

About