China
Africa
Explore chapters and articles related to this topic
Identifying Non-linearity in Construction Workers' Personality Safety Behaviour Predictive Relationship Using Neural Network and Linear Regression Modelling
Published in M.Z. Naser, Leveraging Artificial Intelligence in Engineering, Management, and Safety of Infrastructure, 2023
Yifan Gao, Vicente A. González, Wing Tak Yiu, Guillermo Cabrera-Guerrero
In order to enable the evaluation of network performance, the whole dataset is usually divided into three groups: the training dataset, the validating dataset, and the testing dataset (Ramkumar et al., 2019). The training dataset is used to adjust the network parameters (i.e., weights and biases) to best map the input-output relationships (Klaassen et al., 2016). The validating dataset is used to guarantee that the network is not overfitted while training (Manikandan and Subha, 2016). Overfitting refers to the phenomenon that the learning system overly adapts to the training dataset and even the noise in data, which affects the prediction accuracy of the trained model for new input (Ramkumar et al., 2019). As noted in Ramkumar et al. (2019), it is possible for a network to overfit the training dataset when the MSE value for the validating dataset starts to increase, and the training will be stopped at this point to prevent the network from overfitting. Finally, the testing dataset is established to evaluate the performance of the network after its development (Tümer and Edebali, 2019).
Lung Tumor Segmentation Using a 3D Densely Connected Convolutional Neural Network
Published in Mohan Lal Kolhe, Kailash J. Karande, Sampat G. Deshmukh, Artificial Intelligence, Internet of Things (IoT) and Smart Materials for Energy Applications, 2023
Shweta Tyagi, Sanjay N. Talbar
Out of total 422 CT scans, 330 scans are used for training, 42 scans for validation and 50 scans for testing purpose. Cross-validation by providing the validation data during training helps to control the overfitting problem. Overfitting is a situation where the training accuracy is much larger than the validation accuracy. This is the case when the network learns the training data very well but fails to generalize on validation data. Therefore, rather than testing the overfitted trained model each time, it is better to provide validation data during training, so that if validation accuracy is not improving, the training can be stopped and the model can be improved accordingly. It will save time because training the model each time from scratch till the end is very time-consuming process, so if it is known that the validation accuracy is not increasing, then training can be stopped and then model can be modified by using hyperparameter tuning or in some other way to remove the overfitting or underfitting problem.
Cancer Diagnosis from Histopathology Images Using Deep Learning: A Review
Published in Ranjeet Kumar Rout, Saiyed Umer, Sabha Sheikh, Amrit Lal Sangal, Artificial Intelligence Technologies for Computational Biology, 2023
Vijaya Gajanan Buddhavarapu, J. Angel Arul Jothi
Studies have shown that deeper the network, better the learning ability of the network. However, networks that contain several number of hidden layers increase the complexity of operations performed and require more computational power and large volume of data. DL models require large number of training samples to generalize. Otherwise, the models may overfit and do not perform well on unseen data. Additionally, there is another reason why DL models overfit: Due to the layers of abstraction, DL models may learn to model outliers and rare dependencies. A few methods that are used to combat overfitting include regularization such as L1 regularization, L2 regularization and dropout regularization. Depending upon the architecture, models may be prone to vanishing and exploding gradients. Therefore, DL does come with its own caveat: It requires ample amounts of computational resources and data to unlock its full potential.
Highlighting the present state of biomechanics in shoe research (2000–2023)
Published in Footwear Science, 2023
Benno M. Nigg, Sandro Nigg, Fabian Hoitz, Ashna Subramanium, Jordyn Vienneau, John William Wannop, Arash Khassetarash, Shahab Alizadeh, Emily Matijevich, Eric C. Honert, W. Brent Edwards, Maurice Mohr
The popularity of machine learning algorithms has also been met with valid criticism (Bartlett, 2006). One significant challenge that researchers face when using these techniques is the limitation of available datasets (Ferber et al., 2016; Fukuchi et al., 2017). Specifically in footwear science, machine learning algorithms are often applied to datasets with comparatively small sample sizes (N < 20) that may lack diversity, increasing the risk of overfitted models. Overfitting occurs when a model is too complex and is able to memorise the training data rather than learning underlying patterns. This results in biased models with poor generalisation performances on new, unseen data. Furthermore, biased models may lead to inaccurate or misleading conclusions that cannot be applied to the broader population (Halilaj et al., 2018). Bias can be especially problematic because many machine learning models are considered to be ‘black boxes’ (Horst et al., 2019; Hoitz, Tscharner, et al., 2021). The ‘black box’ characteristic refers to the fact that it is often difficult to understand how a model is making its predictions and classifications, and what factors are influencing the model’s decisions. Typically, a machine learning model is represented by a set of complex mathematical equations that prevent a straightforward interpretation of the relationship between the model’s inputs and outputs.
Bridging the lab-to-field gap using machine learning: a narrative review
Published in Sports Biomechanics, 2023
One major problem of data-driven approaches is that they are optimised for the data they have been trained on. This is particularly a problem when ML is used for small datasets. If the capacity of a ML model is too large and the dataset is too small, the model is only suitable for the specific patterns within the training data and cannot find patterns in new data; it overfits its training data. If the capacity of the model is too low, it cannot represent the relevant features in the training data; it underfits its training data. To avoid underfitting, the capacity of the model needs to be enlarged using more neurons or layers, which increases the likelihood of overfitting. To prevent a model from overfitting, different regularisation strategies have been proposed. The simplest regularisation strategy to prevent the model from overfitting is early stopping of the training procedure. A more sophisticated, regularly employed regularisation strategy is dropout. Dropout means that individual neurons of a network are randomly switched off during the training process to prevent these neurons from specialising in specific features (Goodfellow et al., 1997).
A Novel Data Augmentation Convolutional Neural Network for Detecting Malaria Parasite in Blood Smear Images
Published in Applied Artificial Intelligence, 2022
David Opeoluwa Oyewola, Emmanuel Gbenga Dada, Sanjay Misra, Robertas Damaševičius
Despite these promising results, the proposed approach has several drawbacks. To begin with, all deep learning methods have a tendency to overfit the training dataset. Because the purpose of deep learning models is for them to generalize successfully from training data to any data from the problem domain, it is critical for CNN to make predictions on datasets it has never seen before. Overfitting occurs when a model tries to learn too many details from the training data while still allowing for noise. As a result, the model’s performance on unknown or test datasets is unsatisfactory. This can make the network to fail in generalizing the training dataset’s features or patterns. This inhibits people from making broad generalizations. Moreover, gamma correction may not be the ideal strategy for image enhancement in poor lighting circumstances.