Please use this identifier to cite or link to this item:
|Title:||Converting recurrent neural networks to minimal-state deterministic finite automata for deployment on edge devices||Authors:||Hulagadri, Adithya Venkatadri||Keywords:||Engineering::Computer science and engineering||Issue Date:||2022||Publisher:||Nanyang Technological University||Source:||Hulagadri, A. V. (2022). Converting recurrent neural networks to minimal-state deterministic finite automata for deployment on edge devices. Final Year Project (FYP), Nanyang Technological University, Singapore. https://hdl.handle.net/10356/158314||Abstract:||Edge computing for lowlatency internet-of-things (IoT) applications requires more data analysis to occur closer to the data source on edge devices such as micro-controllers, that have limited memory and computational resources. This project’s objective is to outline a generalisable procedure to classify sequential discrete data using recurrent neural networks (RNNs) that are converted into deterministic finite automata (DFAs). Each DFA can always be minimised using Hopcroft’s algorithm to derive an equivalent DFA with the minimum number of states. This leads to a drastic reduction in the memory required to store the model’s parameters. The extraction of the states from the RNN’s memory is achieved using Quantised Binary Networks (QBNs) which are auto-encoders using a binary activation function that forces the RNN to step through discrete memory states. Lower memory requirements also reduce power consumption, enabling longer service life and applications in bio-computing, which is very heat-sensitive. Additionally, the minimum-state DFA can be always be directly implemented as an electrical circuit, leading to simpler hardware for running advanced RNNs on edge devices. This report utilises a sample classification problem with proven quantum memory advantage. A secondary aim of this report is to demonstrate that the optimum classical algorithm identified for the problem can be deduced from popular machine learning methods. The report shows that this project succeeded in doing so, even obtaining the theoretically optimum DFA from the RNN trained using the Torch framework in the python language.||URI:||https://hdl.handle.net/10356/158314||Schools:||School of Computer Science and Engineering||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Student Reports (FYP/IA/PA/PI)|
Files in This Item:
|1.06 MB||Adobe PDF||View/Open|
Updated on May 31, 2023
Updated on May 31, 2023
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.