Image Recognition based on Dynamic Highway Networks

基於動態高速公路類神經網路之圖像識別研究

王隨賢(Sui-Hsien Wang) ； 林威志(Wei-Zhi Lin) ； 黃漢邦(Han-Pang Huang)

《中國機械工程學刊》 42卷1期 (2021/02) Pp. 23-31

摘要

With the development of machine learning technology, more and more complex networks are developed. For those networks, determining the hyperparameters is important so that they can provide the best performance under the structure of neural network. However, more parameters should be decided in complex networks. This paper is focused on developing a structure of neural networks which can tune the width in each layer based on the utility of neurons automatically. In order to realize this function, a new structure of neural network called convolution neural network based dynamic highway network is proposed to deal with image recognition problem. With the self-adjusting method, near optimal structure and few parameters are required for training to achieve the same and even better performance which uses more neurons.

關鍵字

dynamic highway networks ； machine learning ； highway networks

並列摘要

基於深度學習的發展，許多成功的機器學習模型都相對的巨大。並且需要使用多次的實驗來測試出比較適合的類神經網路結構，此步驟不僅僅耗時也可能找出具有冗餘神經元的架構。為此本文旨在發展一套新式類神經架構，稱為動態高速公路類神經網路。其架構具備高速公路的特性能夠在訓練過程中藉著調整閥門數值，使部分輸入資料傳遞至輸出部分。同時此架構也具備動態調整神經元個數的功能，從較小的類神經網路架構在各層逐漸加入神經元，使得整體類神經模型能夠達到最佳性能。並以MNIST與CIFAR10兩種數據驗證此架構之性能。