Pattern Recognition Based On Machine Learning And Its Implementations On Clinical Technologies

In this thesis,the hardware-friendly machine learning strategies for pattern recognitions are explored by the hardware and software interactive development.Then,several novel and improved algorithms are proposed even implemented for the medical data analysis and building artificial intelligence diagnose assistant systems.Through the real-world case-studies,the recognition efficiency(many aspects of costs)is greatly enhanced by employing a specific machine learning accelerating hardware-platform.From the demo of clinical diagnosis,it is found the efforts of this work is practical in real-world applications.Among plenty of machine learning algorithms for pattern recognition,the sup-port vector machine(SVM)is investigated in this work due to the excellent recog-nition performances.The SVM with Gaussian kernel functions,as one of the most popular algorithms,performs on very high recognition precision since it maps the vectors into a infinite dimensional space.However,a huge number of highly di-mensional Gaussian computations are frequently processed.Both of training and recognizing operations lead to a very high computation cost(time,resource,and data space hungry).To solve this problem,this work adapts a special purpose hardware platform Zac-MaLa(Zhang' Analog Calculator for Machine Learning Accelerator)to implement the SVM learning and recognition.Particularly ad-dressing the spec of this hardware,a hardware-friendly gradient-descent-based backward propagation algorithm is proposed.By collaborating with hardware engineers through interaction of general programming language(C++,perl,etc.)and hardware description language(Verilog,HSPICE,etc.),the case study is done and demonstrated.For convenient demo,the image classification is introduced as examples.The naive SVM has been reported for only two-class classifications(or known as "binary","yes-or-no").In most of real-world applications,multiple classes are generally seen.Many researchers attempted to combine the naive SVM and the decision tree to expend the number of classes.Unfortunately,this principle is hardly applied by non-expert of computer scientists,since the system could be complicated.Thus,I proposed a novel support vector domain description(SVDD)algorithm:only one class of learning samples are needed;by shrinking the domain description,the boundary of sample space is tightly obtained.Namely,a free integration of several SVDDs offers the any demands on the number of classes,even to infinite.Being noticed,these efforts are only feasible and available via the specific hardware mentioned above.In this project,the image processing is exampled again.Thanks to the extremely high speed of hardware,a "constant-scale on-line learning" strategy is proposed,which is distinguished from the conventional "in-cremental on-line learning".In this algorithm,a small scale of learning is done initially;then,the trained model containing SVs is used to recognize an un-labeled vector;after labeling the new vector,it substitutes one of most insignificant SV;the learning is processed again till infinite.In this manner,the learning and recognition is alternatively processed along usage to keep a constant and smal-1 SV scale.This algorithm is helpful to limit the data explosion and increase the learning speed.Moreover,it offers a real-world practical solution for clinical applications.In order to demonstrate how the machine learning supports medical science and engineering,the breast-cancer and heart disease database are introduced and analyzed as case studies.By implementing SVM,the diagnosis is precisely made from a large disease-case data-set.However,when the number of samples decreas-es,miss-diagnosis appears frequently(considering the Chinese clinical situation,very poor database is available).This thesis also presents how to utilize the on-line learning strategy to improve the diagnosis.Since the special hardware in this work is unavailable for most of researchers,I also disclose a toy-example through C++.The rest of this thesis is organized as follows:Chapter 1 introduce the back-ground of machine learning and pattern recognition,and briefly review the their impact on medical science;in Chapter 2,I illustrate SVM principle and the hardware-friendly scheme along with the specific hardware Zac-MaLa,including the case study of image classification;a support vector domain description algo-rithm is explained in Chapter 3;Chapter 4 presents the constant scale on-line learning strategy with case study;in Chapter 5,the clinical application of the proposed systems are demonstrated;finally,conclusions are made in Chapter 6.