深度學習基於訓練數據之技術發展趨勢 : 以專利分析方法探討

Abstract

隨著人工智慧的快速發展，深度學習之神經網絡技術以已成為現今全球技術發展的重點之一，並將其技術運用在各產業領域中。本研究旨在探討深度學習中不同神經網路的技術發展趨勢與應用領域，並透過專利檢索與分析方法來評估其發展趨勢和影響力。通過TIPO全球專利檢索系統資料庫中大量專利數據的收集和分析，探討神經網路技術的歷年專利件數、領先國家別、領先公司別、技術發展現況等，透過專利檢索與技術生命週期分析方法，可以深入了解深度學習技術的應用範圍和為未來發展動向，為未來的研究和產業應用提供價值。總而來說，本研究旨通過專利分析方法深入探討深度學習基於訓練數據之神經網路與其八項神經網絡技術包含循環神經網絡 (Recurrent neural network, RNN) 、卷積神經網絡 (Convolutional Neural Network, CNN) 、生成對抗網絡 (Generative Adversarial Network, GAN) 、時序視覺網絡 (Temporal Segment Networks, TSN) 、自動編碼器 (Autoencoder, AE) 、深度置信網絡 (Deep Belief Network, DBN) 、深度轉移網絡 (Deep Transformation Networks, DTN) 、深度資訊最大化網絡 (Deep InfoMax, DIM)，為相關領域的研與應用提供一定程度的參考依據。With the rapid development of artificial intelligence, deep learning, which involves neural network technologies, has become one of the focal points of global technological advancement. Its applications span across various industries. This study aims to explore the development trends and application domains of different neural network technologies within deep learning. Through patent retrieval and analysis methods, we assess their trends and impacts. By collecting and analyzing a vast amount of patent data from the Global Patent Search System (GPSS) provided by the Taiwan Intellectual Property Office (TIPO), we investigate the yearly number of patents, leading countries, leading companies, and current status of neural network technologies, including recurrent neural networks (RNN), convolutional neural networks (CNN), generative adversarial networks (GAN), temporal segment networks (TSN), autoencoders (AE), deep belief networks (DBN), deep transformation networks (DTN), and deep infomax (DIM). Through patent retrieval and technology lifecycle analysis methods, we gain insights into the application scope and future development trends of deep learning technology, providing valuable references for future research and industrial applications. Overall, this study aims to provide a certain level of reference for research and applications in related fields through patent analysis methods, focusing on deep learning based on training data and eight neural network technologies, including RNN, CNN, GAN, TSN, AE, DBN, DTN, and DIM.