臺灣博碩士論文加值系統

現行藥物開發的動物實驗多數以小鼠做為實驗對象，由於無法透過語言溝通來了解小鼠的感受，目前普遍用以觀察小鼠狀態的方法是透過臉部表情和行為模式。近年來由於機器學習的蓬勃發展，相關的辨識技術與應用獲得大幅度的進展，其中也包含了小鼠臉部及行為判別。然而，監督式的機器學習演算法在分析小鼠表情及動作判別的應用上會遇到兩個主要的挑戰：第一個是需要大量的資料集來減少小鼠間的個體差異對機器學習效能的影響；第二個是由於小鼠的狀態都必須透過第三方觀察得知，因此如何驗證資料標記的可靠性是一大問題，然而準確可靠的資料標記對於機器學習模型的訓練是非常基本的要求。特別是在小鼠的痠、痛分析的問題上，人類無法與小鼠透過言語溝通判斷正確的資料標記。
本論文將著重於訓練樣本與特徵於不同分類架構的效應分析，用以克服上述的兩個挑戰。比較的架構包含產生訓練樣本的方法、小鼠的行為特徵萃取、影像上的時間特徵以及不同類別的機器學習模型，透過實驗分析這些不同架構於相同資料集所產生出來的結果、驗證是否符合預期。產生出來的結果主要是透過觀察準確率、誤報率，和檢測率去觀察不同系統於少量資料與未曾包含於訓練過程的新個體表現。本篇論文並提出了一個非監督式的方法與其他架構作結合，嘗試減少過擬合的現象還有資料標記不夠精確的影響。

Mice have been an important reference for the drug test. Currently, it is only possible to collect the feedbacks from mice through non-verbal methods. The most common ways to observe the status of mice are through the facial expressions or behaviors. Due to recent advance of machine learning, its utilities have been demonstrated in many applications. One of them is analysis in the facial expressions or behaviors of mice. However, there are usually two major challenges when applying supervised machine learning algorithms. The first one is individual variation on facial expression or behaviors, which might require a huge amount of data set to overcome. The second one is how to obtain reliable labels, which are fundamental to train a robust machine learning model. Especially in the research of sng and pain of mice, human beings are unable to label the data set by verbal communications.
This thesis aimed on the analysis of the effects of training samples and features on different classification architectures to overcome the above two challenges. The compared architectures included selection of training samples, feature extraction of mouse behavior, and temporal resolution of video frames combining with different types of machine learning models. These architectures were analyzed on the same dataset through the experiments. The accuracy, false alarm and detection rate were used to evaluate their performance on a small data set and the individuals excluded from the training procedure of the models. This thesis also proposed an unsupervised approach that can be embedded into these architectures in order to reduce the impact of overfitting and non-reliable labels.

摘 要 i
ABSTRACT iii
Table of Contents v
List of Tables vii
List of Figures viii
Chapter 1 INTRODUCTION 1
1.1 Motivation and Background 1
1.2 Objective 2
1.3 Organization 2
Chapter 2 OVERVIEW 3
2.1 DeepLabCut 3
2.2 Motion analysis methods 3
2.2.1 B-SOiD 3
2.2.2 Repetition estimation 4
2.3 Pain detection through pose features 5
2.4 Classification models 6
2.4.1 Support vector machine (SVM) 6
2.4.2 Random Forest (RF) 6
2.4.3 Multiple layer perceptron (MLP) 7
2.4.4 Long short-term memory (LSTM) 7
Chapter 3 METHOLOGY 9
3.1 Fundamental Architecture 10
3.2 Feature extraction 10
3.2.1 Frame-based features 11
3.2.2 Segment-based features 11
3.3 Clustering step embedding 15
3.4 Models 16
Chapter 4 EXPERIMENT STUDIES 18
4.1 Datasets 18
4.2 Metrics 19
4.3 Experimental setup 20
4.4 Analysis 21
4.4.1 Effects of the frame/segment features 21
4.4.2 Effects of clustering step 22
4.4.3 Effects of CWT features 23
4.4.4 Effects of sample overlapping 24
4.4.5 Effects of different models 25
4.5 Overall comparison 26
4.6 Analysis of Blind-testing set 27
4.7 Feature importance 29
Chapter 5 CONCLUSION 32
References 33
Appendix A – All architectures and experimental results 34

[1]Mathis, Alexander, et al. “DeepLabCut: markerless pose estimation of user-defined body parts with deep learning.” Nature neuroscience 21.9 (2018): 1281-1289.
[2]Hsu, A.I., Yttri, E.A. “B-SOiD, an open-source unsupervised algorithm for identification and fast prediction of behaviors.” Nat Commun 12, 5188 (2021).
[3]Runia, T.F.H., Snoek, C.G.M. & Smeulders, A.W.M. “Repetition Estimation.” Int J Comput Vis 127, 1361–1383 (2019).
[4]Torrence, C., & Compo, G. P. (1998). A practical guide to wavelet analysis. Bulletin of the American Meteorological Society, 79(1), 61–78.
[5]KUO, TING-KAI. “Pain detection through mouse pose features from depth images” https://hdl.handle.net/11296/ju8p5a
[6]Cortes, Corinna, and Vladimir Vapnik. "Support-vector networks." Machine learning 20.3 (1995): 273-297.Sad.
[7]Tin Kam Ho, “Random decision forests,” Proceedings of 3rd International Conference on Document Analysis and Recognition. Montreal, QC, 14–16 August 1995. pp. 278–282.
[8]J. R. Quinlan, C4.5: Programs for Machine Learning. Elsevier, 2014.
[9]S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997.
[10]C. Segalin, J. Williams, T. Karigo, M. Hui, M. Zelikowsky, J. J. Sun, P. Perona, D. J. Anderson, and A. Kennedy, “The Mouse Action Recognition System (MARS) software pipeline for automated analysis of social behaviors in mice,” eLife, vol. 10, 2021.
[11]Breiman, Leo.“Random Forests.” Machine Learning 45 (1). Springer: 5-32 (2001).