Year: 2020
Type: other
Abstract: The family of machine learning architecture called Recurrent Neural Network (RNN) has improved the processing of temporal dynamic behavior. Unlike other Artificial Neural Networks, they have the capacity to use an internal state to memorize information during the processing of a sequence of inputs. Recurrent neural networks use a complex architecture to retain useful information from the input and reuse it later even if the delay between two inputs is important. These networks and algorithms are mainly used to process sequences, for example, unsegmented gesture recognition, connected handwriting recognition, speech recognition, or activity detection. The goal of this internship is to study and propose new approaches for RNNs that can adapt to changing computational resources, changing constraints, and resource-limited edge devices while maintaining the best trade-off between performance and computation cost. In order to predict unique, multiple values or labels from an input sequence we review different types of recurrent neural network architectures for supervised machine learning. We first review the state of the art in flexible RNNs in order to understand the existing methods. Then we describe our novel RNN architecture that can limit computation over a time lap. Finally, we evaluate our approach on three data set (HAR, Adding Task, and Mnist). These experimentations show that our approach offers accuracy/ computational trade-offs similar to the state-of-the-art, and, unlike these methods, strictly upper bounds the computational cost over a time lap.
Source: Doria (University of Helsinki)
Author: Tsiry Mayet
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3110025856
