China
Africa
Explore chapters and articles related to this topic
Deep Learning
Published in Subasish Das, Artificial Intelligence in Highway Safety, 2023
Recursive neural networks (not the same as RNNs) are non-linear adaptive models utilized to analyze data of variable length that feed the state of the network back into itself in a loop and that are capable of processing data structure inputs and are mostly suited for sentence and image deconstruction. Their architecture permits users to both determine the components of input data and quantitatively assess their relationships through a binary tree structure and a shared-weight matrix, which permits the recursive neural network to extrapolate from variable-length sequences of words and images. Recursive networks also have the advantage that for a sequence of length n, the depth (which is given as the number of compositions of nonlinear operations) can be taken from n to log(n), which permits efficient capturing of long-term dependencies. Recursive neural networks typically have a top-down propagation method and a bottom-up feed-forward method. Both these mechanisms make up the propagation through a common structure found in the majority of recursive networks.
Analyzing the Suitability of Deep Learning Algorithms in Edge Computing
Published in Sam Goundar, Archana Purwar, Ajmer Singh, Applications of Artificial Intelligence, Big Data and Internet of Things in Sustainable Development, 2023
In Recursive Neural Network the computation performed in one layer is used by the other layer. So, each layer is dependent on the last computation. For this RNN uses inbuilt memory. The samples given to RNN have interdependencies, and it remembers all the previous steps.
MS-TR: A Morphologically enriched sentiment Treebank and recursive deep models for compositional semantics in Turkish
Published in Cogent Engineering, 2021
Sultan Zeybek, Ebubekir Koç, Aydın Seçer
Recursive Neural Network (Tree-RNN) is a tree-structured model based on composing words over nested hierarchical structure in sentences. The neural network function recursively merges words to construct noun phrases until representing the entire sentence. Tree-RNNs have an extraordinary ability for mapping of phrases in a semantic space (Socher et al., 2010).
Simultaneous identification of thermophysical properties of semitransparent media using an artificial neural network trained by a 2-D axisymmetric direct model
Published in Numerical Heat Transfer, Part A: Applications, 2020
Yang Liu, Yann Billaud, Didier Saury, Denis Lemonnier
The choice of the ANN. The recursive neural network, which is a type of ANN design to apply the same set of weights recursively over a structured input, may be a good candidate given its reputation to deal with sequential data and thus temporal series.
A Novel Semantic-Enhanced Text Graph Representation Learning Approach through Transformer Paradigm
Published in Cybernetics and Systems, 2023
Along with recent raises of deep learning (Gasmi et al. 2021), advanced neural network architectures (Guo, Li, and Zhan 2021), such as sequence-to-sequence (seq2seq) (Sutskever, Vinyals, and Le 2014; Bahdanau, Cho, and Bengio 2015), attention mechanism and transformer (Ashish et al. 2017; Devlin et al. 2019) have drawn a lot of interests from researchers and organizations recent years. Staring from the well-known large-scale corpus-based word representation learning approach, called as Word2Vec (Mikolov et al. 2013) which presented remarkable improvements in textual representation learning. Inherited from success of Word2Vec, other local contextual word embedding techniques, e.g., GloVe (Pennington, Socher, and Manning 2014), fastText (Mikolov et al. 2018), and Doc2Vec (Le and Mikolov 2014) have been proposed. These word embedding approaches and their related learning paradigms (Szymański and Kawalec 2019) have showed significant increases in accuracy performance of multiple NLP’s tasks, including classification (Onan 2019; 2021). However, recent word embedding-based textual representation learning techniques still encounter limitations related to the capability of capturing rich-semantic sequential and contextual representations of texts. Thus, multiple complex recurrent neural network (RNN)-based architectures, (e.g. Tree-LSTM (Tai, Socher, and Manning 2015), RT-LSTM (Zhu, Sobihani, and Guo 2015), MST-LSTM (Liu et al. 2015), etc.) been proposed to effectively capture both sequential and syntactic information of consecutive word sequences for leveraging the performance of text classification. On the other side, there are proposed convolutional neural network (CNN)-based textual representation learning techniques which take the advantages of multilayered CNN-based architecture to deeply learn and characterize the distinctive pattern-based features of texts upon the learnt sentence/word embedding matrices. There are well-known works (e.g., Dynamic-CNN (Blunsom, Grefenstette, and Kalchbrenner 2014), VDCNN (Conneau et al. 2017), etc.) of applying CNN-based text representation learning mechanism which demonstrated effectiveness in multiple NLP-based task-driven training purposes. However, both RNN- and CNN-based textual representation approach still suffers limitations of unable to capturing the global long-range dependent relationships of words during the textual embedding process. Thus, they still cannot reach the highest performance within complex textual corpora in which their documents carry long-distance non-sequential-ordered relationships between consecutive words. Thus, a new research direction is necessary to meet this challenge.