China
Africa
Explore chapters and articles related to this topic
Theoretical Paradigms in Cognitive Science and in Theoretical Neurophilosophy
Published in Harald Maurer, Cognitive Science, 2021
The "Adaptive Resonance Theory (ART)",185 put forward by the American mathematicians and neuroinformaticians Steven Grossberg and Gail A. Carpenter, also uses a self-organized learning algorithm within the framework of unsupervised competitive learning. They have developed an entire class of highly plausible models (neurobiologically and cognitive psychologically speaking), designed to solve the "stability plasticity dilemma"186 of artificial neural networks. This dilemma, also known as the classification problem, is to account for how vectorial information (as a sequence of input patterns to be learned) is independently integrated into a (synchronous) statistical prototype, such that the already learned associations of the network can be adapted toward new input patterns ("plasticity") without causing these old associations to be modified too much ("catastrophic forgetting" (Grossberg and Versace 2008)). This ensures the preservation of the patterns once learned ("stability"). In other words, the classification problem leads to the fundamental question: "How can new associations be learned in a neural network without forgetting old associations?" (Zell 2003).
Artificial neural network models of sports motions
Published in Youlian Hong, Roger Bartlett, Routledge Handbook of Biomechanics and Human Movement Science, 2008
W.I. Schöllhorn, J.M. Jäger, D. Janssen
Once the network has become tuned to the statistical regularities of the input data, it develops the ability to form internal representations for encoding features of the input and thereby generates classes automatically (Becker, 1991). Learning vector quantization (LVQ) is a supervised learning extension of the Kohonen network methods (Kohonen, 2001). Another form of unsupervised learning network is based on the adaptive resonance theory (ART) developed by Carpenter and Grossberg (1987). ART-1 Network, as introduced by Grossberg (Grossberg, 1976 b; Grossberg, 1976 a) is a self-organizing and self-stabilizing vector classifier where the environment modulates the learning process and thereby performs an unsupervised teaching role. Input vectors are added to existing feature clusters and are, therefore, able to learn new examples after the initial training has been completed. Finally, the approach of dynamically controlled networks (DyCoN) (Perl, 2000) allows for dynamically self-adapting the training process to continuously changing situations. One advantage of DyCoNs is seen in their trainability. Once a SOM is trained it can be applied quite successfully for interpolation. If data occur that are outside the trained area the SOM has to be trained again from the beginning. In contrast to this a trained DyCoN can be trained again with new data sets without starting from the beginning. This provides an additional advantage, since it allows to simulate qualitatively learning and adaptation processes as well (Perl and Baca, 2003). Most recently, Dynamically Controlled Networks (Perl, 2002) are combined with Neural Gases (Martinetz and Schulten, 1991; Martinetz and Schulten, 1994; Fritzke, 1995) to DyConG-models (Memmert and Perl, 2005).
A machine-learning method for improving crash injury severity analysis: a case study of work zone crashes in Cairo, Egypt
Published in International Journal of Injury Control and Safety Promotion, 2020
Mahama Yahaya, Wenbo Fan, Chuanyun Fu, Xiang Li, Yue Su, Xinguo Jiang
Many past studies have investigated the crash injury severity with potential risk factors, using machine learning classification methodologies. Typically, classifiers are made to identify which set of categories a new observation belongs to, based on the training data containing observations with known class membership. For example, Abdelwahab et al. (Abdelwahab & Abdel-Aty, 2001) applied the multilayer perceptron and fuzzy adaptive resonance theory based on a neural network to classify the injury severity of drivers with vehicular, human, roadway, and environmental factors. Similarly, Chang and Wang (Chang & Wang, 2006) applied the classification and regression tree (CART) to study crash data from police records collected in Taiwan. Further, (C. Chen et al., 2015) proposed a hybrid technique that combined the Bayesian network and multinomial logit to classify the driver injury severity, using New Mexico crash data. Recently, Jeong et al. (Jeong et al., 2018) applied different classification models including Logistic regression, Decision tree, Neural network, Gradient boosting model, and Naïve Bayes classifier to model motor vehicle crash injury severity in imbalanced crash data. Similarly, (Vilaça et al., 2019) applied classifiers-decision tree and logistic regression, in the context of rare event modelling, to assess the injury severity risk of vulnerable road users. Through the extensive literature review herein, it is obvious that the application of machine learning classifiers for the traffic crash severity studies has grown rapidly over the last few years.