Explore chapters and articles related to this topic
Wind Power Forecasting via Deep Learning Methods
Published in Jacqueline A. Stagner, David S-K. Ting, Green Energy and Infrastructure, 2020
LSTM and GRU both are RNN methods. GRU is different from the LSTM according to utilizing gating information to prevent vanishing gradient problem. GRU has gating units that control the data flow. Furthermore, GRU does not have a specialized memory cell. Figure 7.2 discusses the gates of LSTM and GRU. i corresponds to the input gate, f shows the forget gate, o shows the output gate, and c and c˜ are the memory cell and the cell content. GRU includes reset and update gates, activation, and candidate activation. r and z represent the reset and update gates, whereas h and h˜ denote the activation and the candidate activation, respectively (Chung, Gulcehre, Cho, & Bengio, 2014).
Deep Learning in Transportation Cyber-Physical Systems
Published in M.Z. Naser, Leveraging Artificial Intelligence in Engineering, Management, and Safety of Infrastructure, 2023
Zadid Khan, Sakib Mahmud Khan, Mizanur Rahman, Mhafuzul Islam, Mashrur Chowdhury
A modified version of the LSTM neuron was created to reduce its complexity, which is known as a gated recurrent unit (GRU). A new “update gate” is created in GRU, which combines the input gate and forget gate. The cell state and hidden states are combined for faster operation. Equations 14.11–14.13 describe the mathematical operations behind GRU, where the intermediate states are zt and rt.
A spatial-temporal feature fusion network for order remaining completion time prediction in discrete manufacturing workshop
Published in International Journal of Production Research, 2023
Shengbo Wang, Yu Guo, Shaohua Huang, Daoyuan Liu, Pengzhou Tang, Litong Zhang
The GRU consists of an update gate, a reset gate, a current memory, and a final memory. The reset gate determines how the new input is combined with the previous memory. The update gate determines the proportion of the previous memory occupied. The calculation steps are as follows: where represents two matrices are multiplied by the elements; represents the data at the current time; denotes the output of the previous network unit; represents the output of update gate; denotes the output of reset gate; denotes candidate hidden state; represents the putout of the current unit; , , , , , denote the weights to be learned in the network.
Multimodal face shape detection based on human temperament with hybrid feature fusion and Inception V3 extraction model
Published in Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2023
Srinivas Adapa, Vamsidhar Enireddy
A GRU is considered one of the novel memory cells used in various real-time applications due to its effectiveness in detecting and classification problems. It is an extension of traditional long short termshort-term memory (LSTM). Compared to LSTM, GRU is composed of two gates, such as update and reset, whereas LSTM is composed of three gates, such as input, forget, and output. The reset gate in the GRU has the responsibility of combining the new input with the previous memory. The update gate has the responsibility of keeping the required previous states. GRU is one of the advanced versions of RNN (recurrent neural network) consistsconsisting of a gating mechanism. It can prevent the vanishing gradient issue that occurs in traditional RNN. GRU is closer to the traditional LSTM model but has few training parameters. GRU is faster than LSTM, with less memory consumption and training time. Without the help of any control, the GRU can expose the entire hidden states. In GRU, for human character classification based on facial features, the softmax layer is used as the final layer for both multi-class and binary classification. This layer works fast and provides solutions with a probabilistic description. In the proposed work, the classifier works based on the face region and classifies the human character. The softmax layer is considered the classification layer and is mathematically represented as,
Prediction of tunnel mechanical behaviour using multi-task deep learning under the external condition
Published in Georisk: Assessment and Management of Risk for Engineered Systems and Geohazards, 2023
Bowen Du, Tao Zou, Junchen Ye, Xuyan Tan, Ke Cheng, Weizhong Chen
A recurrent neural network (RNN) is a kind of neural network designed for sequence data. It captures the periodic patterns in the sequence and fuses the historical and current information. However, RNN fails to capture the long-term dependencies and some problems such as gradient vanishing appear when modelling long sequences of series. Therefore, some variants of RNN such as GRU and LSTM are proposed for avoiding these phenomena. Different from RNN, GRU uses gate mechanisms to control the information between earlier parts and current input. In GRU, it provides two gate mechanisms to save and discard information at each time step, namely the Update gate and the Reset gate. Both of these gates are trained separately to capture the available information. The process of GRU is formalised as where Wr,Wz,Wh,br,bz,bh are learnable parameters, ⊙ denotes the Hadamard product and σ and tanh are the activation functions. To capture the spatial dependencies when learning temporal patterns, we replace the transform for hidden states and input data in the GRU module as a graph convolution network. We first update the representations of hidden states and current features for all nodes in the network through multiple layers of GCN. Then the updated embeddings are put into two gate mechanisms in GRU for learning temporal dependencies. The process of AGGRU unit is defined as where GCN(·) represents the multiple layers of the graph convolution network shown in Equation (4), is the updated state of all nodes in the network.