Professional Documents
Culture Documents
Long Short-Term Memory Survey Paper
Long Short-Term Memory Survey Paper
Long Short-Term Memory Survey Paper
A survey on “Long_Short-Term_Memory”
Tushar sharma
19ECTCS074
Computer Science and Engineering
University College of Engineering and Technology, Bikaner
In order to model two sequence datasets, this paper In this paper, we evaluate the effectiveness of a
carried an empirical comparison between the recent approach that simplifies the standard LSTM
traditional LSTM and three new simplified architecture. We compare three simplified LSTM
variations that were created by removing input variants with the standard LSTM on two datasets.
signals, bias, and hidden unit signals from Our results show that the simplified LSTMs is
individual gates. capable of achieving comparable performance to
the standard LSTM.
I. INTRODUCTION
Fundamentals of LSTM
RNN a type of deep complex network, have shown
great power in sequence modeling tasks such as LSTM is a type of recurrent neural network (RNN)
language translation, speech recognition, and image architecture that has been focused on the ability to
captioning. Unlike feedforward neural_networks capture and generate long-term dependence in
such as CNN, RNNs have cyclic connections and sequential data Temporal dependence hinders
hidden states based on elapsed time steps but RNNs learning
face challenges in training due to vanishing-burst-
prone problems, limiting their ability to capture The LSTM network consists memory cell, which
long-term dependencies. can hold long sequences in the network.
To overcome these challenges, researchers The memory cell contains an auto-loop that stores
developed a modified RNN algorithm called its value from the previous time step, ensuring that
LSTM. The LSTM consists of a memory cell the reference can be stored and used in time
capable of maintaining information over time and a
gating mechanism with input, output and forget Gating devices in LSTM networks control flow of
gates. This algorithm effectively captures and information into memory cell. There are of three
exploits the long-term dependence without the types:
training difficulties of traditional RNNs.
In LSTM, several improvements and variants have 1. Input Gate: It controls how new information is
been proposed. Peephole connections were added to added in memory cell at current time step. It takes
enable precise timing of outputs, recurrent and non- into account current input and previously stored
state, and by processing it determines the
appropriate information for updating the memory dependencies. Here are some commonly used
cell. LSTM architectures and variants:
Applications of LSTM_Models
4. www.sciencedirect.com
5. www.researchgate.net