This work demonstrates the viability of the ballistocardiogram (BCG) signal derived from a head-worn device as a biometric modality for authentication. The BCG signal is the measure of an individual's body acceleration as a result of the heart's ejection of blood. It is a characterization of an individual's cardiac cycle and can be derived non-invasively from the measurement of subtle movements of a person's extremities. Through the use of accelerometer and gyroscope sensors on a Smart Eyewear (SEW) device, derived BCG signals are used to train a convolutional neural network (CNN) as an authentication model, which is personalized for each wearer. This system is evaluated using data from 12 subjects, showing that this approach has an equal error rate of 3.5% immediately after training, and only marginally degrades to 13% after about 2 months, in the worst case. We also explore the use of our authentication approach for individuals with severe motor disabilities, and observe that the results fall only slightly short of those of the larger population, with immediate EER values at 11.2% before rising to 21.6%, again in the worst case.. Overall, we demonstrate that this model presents a longitudinally-viable authentication solution for passive biometric authentication.

Early time series classification is the task of predicting the class label of a time series before it is observed in its entirety. In time-sensitive domains where information is collected over time it is worth sacrificing some classification accuracy in favor of earlier predictions, ideally early enough for actions to be taken. However, since accuracy and earliness are contradictory objectives, a solution to this problem must find a task-dependent trade-off. There are two common state-of-the-art methods. The first involves an analyst selecting a timestep at which all predictions must be made. This does not capture earliness on a case-by-case basis, so if the selecting timestep is too early, all later signals are missed, and if a signal happens early, the classifier still waits to generate a prediction. The second method is the exhaustive search for signals, which encodes no timing information and is not scalable to high dimensions or long time series. We design the first early classification model called EARLIEST to tackle this multi-objective optimization problem, jointly learning (1) to decide at which time step to halt and generate predictions and (2) how to classify the time series. Each of these is learned based on the task and data features. We achieve an analyst-controlled balance between the goals of earliness and accuracy by pairing a recurrent neural network that learns to classify time series as a supervised learning task with a stochastic controller network that learns a halting-policy as a reinforcement learning task. The halting-policy dictates sequential decisions, one per timestep, of whether or not to halt the recurrent neural network and classify the time series early. This pairing of networks optimizes a global objective function that incorporates both earliness and accuracy. We validate our method via critical clinical prediction tasks in the MIMIC III database from the Beth Israel Deaconess Medical Center along with another publicly available time series classification dataset. We show that EARLIEST out-performs two state-of-the-art LSTM-based early classification methods. Additionally, we dig deeper into our model's performance using a synthetic dataset which shows that EARLIEST learns to halt when it observes signals without having explicit access to signal locations. The contributions of this work are three-fold. First, our method is the first neural network-based solution to early classification of time series, bringing the recent successes of deep learning to this problem. Second, we present the first reinforcement-learning based solution to the unsupervised nature of early classification, learning the underlying distributions of signals without access to this information through trial and error. Third, we propose the first joint-optimization of earliness and accuracy, allowing learning of complex relationships between these contradictory goals.

by John Pui-Fai Sum. / Thesis (Ph.D.)--Chinese University of Hong Kong, 1998. / Includes bibliographical references (p. 155-[163]). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Mode of access: World Wide Web.

Artificial recurrent neural networks (RNNs) are powerful models for understanding and modeling dynamic computation in neural circuits. As such, RNNs that have been constructed to perform tasks analogous to typical behaviors studied in systems neuroscience are useful tools for understanding the biophysical mechanisms that mediate those behaviors. There has been significant progress in recent years developing gradient-based learning methods to construct RNNs. However, the majority of this progress has been restricted to network models that transmit information through continuous state variables since these methods require the input-output function of individual neuronal units to be differentiable. Overwhelmingly, biological neurons transmit information by discrete action potentials. Spiking model neurons are not differentiable and thus gradient-based methods for training neural networks cannot be applied to them. This work focuses on the development of supervised learning methods for RNNs that do not require the computation of derivatives. Because the methods we develop do not rely on the differentiability of the neural units, we can use them to construct realistic RNNs of spiking model neurons that perform a variety of benchmark tasks, and also to build networks trained directly from experimental data. Surprisingly, spiking networks trained with these non-gradient methods do not require significantly more neural units to perform tasks than their continuous-variable model counterparts. The crux of the method draws a direct correspondence between the dynamical variables of more abstract continuous-variable RNNs and spiking network models. The relationship between these two commonly used model classes has historically been unclear and, by resolving many of these issues, we offer a perspective on the appropriate use and interpretation of continuous-variable models as they relate to understanding network computation in biological neural circuits. Although the main advantage of these methods is their ability to construct realistic spiking network models, they can equally well be applied to continuous-variable network models. An example is the construction of continuous-variable RNNs that perform tasks for which they provide performance and computational cost competitive with those of traditional methods that compute derivatives and outperform previous non-gradient-based network training approaches. Collectively, this thesis presents efficient methods for constructing realistic neural network models that can be used to understand computation in biological neural networks and provides a unified perspective on how the dynamic quantities in these models relate to each other and to quantities that can be observed and extracted from experimental recordings of neurons.

Liu Shubao. / Thesis (M.Phil.)--Chinese University of Hong Kong, 2005. / Includes bibliographical references (leaves 90-98). / Abstracts in English and Chinese. / Abstract --- p.i / 摘要 --- p.iii / Acknowledgement --- p.iv / Chapter 1 --- Introduction --- p.1 / Chapter 1.1 --- Time-Varying Quadratic Optimization --- p.1 / Chapter 1.2 --- Recurrent Neural Networks --- p.3 / Chapter 1.2.1 --- From Feedforward to Recurrent Networks --- p.3 / Chapter 1.2.2 --- Computational Power and Complexity --- p.6 / Chapter 1.2.3 --- Implementation Issues --- p.7 / Chapter 1.3 --- Thesis Organization --- p.9 / Chapter I --- Theory and Models --- p.11 / Chapter 2 --- Linearly Constrained QP --- p.13 / Chapter 2.1 --- Model Description --- p.14 / Chapter 2.2 --- Convergence Analysis --- p.17 / Chapter 3 --- Quadratically Constrained QP --- p.26 / Chapter 3.1 --- Problem Formulation --- p.26 / Chapter 3.2 --- Model Description --- p.27 / Chapter 3.2.1 --- Model 1 (Dual Model) --- p.28 / Chapter 3.2.2 --- Model 2 (Improved Dual Model) --- p.28 / Chapter II --- Engineering Applications --- p.29 / Chapter 4 --- KWTA Network Circuit Design --- p.31 / Chapter 4.1 --- Introduction --- p.31 / Chapter 4.2 --- Equivalent Reformulation --- p.32 / Chapter 4.3 --- KWTA Network Model --- p.36 / Chapter 4.4 --- Simulation Results --- p.40 / Chapter 4.5 --- Conclusions --- p.40 / Chapter 5 --- Dynamic Control of Manipulators --- p.43 / Chapter 5.1 --- Introduction --- p.43 / Chapter 5.2 --- Problem Formulation --- p.44 / Chapter 5.3 --- Simplified Dual Neural Network --- p.47 / Chapter 5.4 --- Simulation Results --- p.51 / Chapter 5.5 --- Concluding Remarks --- p.55 / Chapter 6 --- Robot Arm Obstacle Avoidance --- p.56 / Chapter 6.1 --- Introduction --- p.56 / Chapter 6.2 --- Obstacle Avoidance Scheme --- p.58 / Chapter 6.2.1 --- Equality Constrained Formulation --- p.58 / Chapter 6.2.2 --- Inequality Constrained Formulation --- p.60 / Chapter 6.3 --- Simplified Dual Neural Network Model --- p.64 / Chapter 6.3.1 --- Existing Approaches --- p.64 / Chapter 6.3.2 --- Model Derivation --- p.65 / Chapter 6.3.3 --- Convergence Analysis --- p.67 / Chapter 6.3.4 --- Model Comparision --- p.69 / Chapter 6.4 --- Simulation Results --- p.70 / Chapter 6.5 --- Concluding Remarks --- p.71 / Chapter 7 --- Multiuser Detection --- p.77 / Chapter 7.1 --- Introduction --- p.77 / Chapter 7.2 --- Problem Formulation --- p.78 / Chapter 7.3 --- Neural Network Architecture --- p.82 / Chapter 7.4 --- Simulation Results --- p.84 / Chapter 8 --- Conclusions and Future Works --- p.88 / Chapter 8.1 --- Concluding Remarks --- p.88 / Chapter 8.2 --- Future Prospects --- p.88 / Bibliography --- p.89

模型預測控制是一種基於模型的先進控制策略，它通過反復優化一個有限時域内的約束優化問題實時求解最優控制信號。作爲一種有效的多變量控制方法，模型預測控制在過程控制、機械人、經濟學等方面取得了巨大的成功。模型預測控制研究與發展的一個關鍵問題在於如何實現高性能非綫性和魯棒預測控制算法。實時優化是一項具有挑戰性的任務，尤其在優化問題時非凸優化的情況下，實時優化變得更爲艱巨。在模型預測控制取得發展的同時，以建立仿腦計算模型為目標的神經網絡研究也取得一些重要突破，尤其是在系統辨識和實時優化方面。神經網絡為解決模型預測控制面臨的瓶頸問題提供了有力的工具。 / 本篇論文重點討論基於神經動力學方法的模型預測控制的設計與分析。論文的主要目標在於設計高性能神經動力學算法進而提高模型預測控制的最優性與計算效率。論文包括兩大部分。第一部分討論如何在不需要求解非凸優化的前提下解決非綫性和魯棒模型預測控制。主要的解決方案是將非相信模型分解為帶有未知項的仿射模型，或將非綫性模型轉換為綫性變參數系統。仿射模型中的未知項通過極限學習機進行建模和數值補償。針對系統中的不塙定干擾，利用極小極大算法和擾動不變集方法獲取控制系統魯棒性。儅需要考慮多個評價指標是，採用目標規劃設計多目標優化算法。論文第一部分提出的設計方法可以將非綫性和魯邦模型預測控制設計為凸優化問題，進而採用神經動力學優化的方法進行實時求解。論文的第二部分設計了針對非凸優化的多神經網絡算法，並在此基礎上提出了模型預測控制算法。多神經網絡算法模型人類頭腦風暴的過程，同時應用多個神經網絡相互協作地進行全局搜索。神經網絡的動態方程指導其進行局部精確搜索，神經網絡之間的信息交換指導全局搜索。實驗結果表明該算法可以高效地獲得非凸優化的全局最優解。基於多神經網絡優化的的模型預測控制算法是一種創新性的高性能控制方法。論文的最後討論了應用模型預測控制解決海洋航行器的運動控制問題。 / Model predictive control (MPC) is an advanced model-based control strategy that generates control signals in real time by optimizing an objective function iteratively over a finite moving prediction horizon, subject to system constraints. As a very effective multivariable control technology, MPC has achieved enormous success in process industries, robotics, and economics. A major challenge of the MPC research and development lies in the realization of high-performance nonlinear and robust MPC algorithms. MPC requires to perform real time dynamic optimization, which is extremely demanding in terms of solution optimality and computational efficiency. The difficulty is significantly amplified when the optimization problem is nonconvex. / In parallel to the development of MPC, research on neural networks has made significant progress, aiming at building brain-like models for modeling complex systems and computing optimal solutions. It is envisioned that the advances in neural network research will play a more important role in the MPC synthesis. This thesis is concentrated on analysis and design of neurodynamic approaches to nonlinear and robust MPC. The primary objective is to improve solution optimality by developing highly efficient neurodynamic optimization methods. / The thesis is comprised of two coherent parts under a unified framework. The first part consists of several neurodynamics-based MPC approaches, aiming at solving nonlinear and robust MPC problems without confronting non-convexity. The nonlinear models are decomposed to input affine models with unknown terms, or transformed to linear parameter varying systems. The unknown terms are learned by using extreme learning machines via supervised learning. Minimax method and disturbance invariant tube method are used to achieve robustness against uncertainties. When multiobjective MPC is considered, goal programming technique is used to deal with multiple objectives. The presented techniques enable MPC to be reformulated as convex programs. Neurodynamic models with global convergence, guaranteed optimality, and low complexity are customized and applied for solving the convex programs in real time. Simulation results are presented to substantiate the effectiveness and to demonstrate the characteristics of proposed approaches. The second part consists of collective neurodynamic optimization approaches, aiming at directly solving the constrained nonconvex optimization problems in MPC. Multiple recurrent neural networks are exploited in framework of particle swarm optimization by emulating the paradigm of brainstorming. Each individual neural network carries out precise constrained local search, and the information exchange among neural networks guides the improvement of the solution quality. Implementation results on benchmark problems are included to show the superiority of the collective neurodynamic optimization approaches. The essence of the collective neurodynamic optimization lies in its global search capability and real time computational efficiency. By using collective neurodynamic optimization, high-performance nonlinear MPC methods can be realized. Finally, the thesis discusses applications of MPC on the motion control of marine vehicles. / Detailed summary in vernacular field only. / Detailed summary in vernacular field only. / Yan, Zheng. / Thesis (Ph.D.) Chinese University of Hong Kong, 2014. / Includes bibliographical references (leaves 186-203). / Abstracts also in Chinese.

Han-bing Ji. / Thesis (Ph.D.)--Chinese University of Hong Kong, 1996. / Includes bibliographical references (p. 160-163). / Microfiche. Ann Arbor, Mich.: UMI, 1998. 2 microfiches ; 11 x 15 cm.

Guo Ping. / "March 2002." / Thesis (Ph.D.)--Chinese University of Hong Kong, 2002. / Includes bibliographical references (p. 166-182). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Mode of access: World Wide Web. / Abstracts in English and Chinese.

During the past two decades, numerous recurrent neural networks (RNNs) have been proposed for solving VIs and related problems. However, first, the theories of many emerging RNNs have not been well founded yet; and their capabilities have been underestimated. Second, these RNNs have limitations in handling some types of problems. Third, it is certainly not true that these RNNs are best choices for solving all problems, and new network models with more favorable characteristics could be devised for solving specific problems. / In the research, the above issues are extensively explored from dynamic system perspective, which leads to the following major contributions. On one hand, many new capabilities of some existing RNNs have been revealed for solving VIs and related problems. On the other hand, several new RNNs have been invented for solving some types of these problems. The contributions are established on the following facts. First, two existing RNNs, called TLPNN and PNN, are found to be capable of solving pseudomonotone VIs and related problems with simple bound constraints. Second, many more stability results are revealed for an existing RNN, called GPNN, for solving GVIs with simple bound constraints, and it is then extended to solve linear VIs (LVIs) and generalized linear VIs (GLVIs) with polyhedron constraints. Third, a new RNN, called IDNN, is proposed for solving a special class of quadratic programming problems which features lower structural complexity compared with existing RNNs. Fourth, some local convergence results of an existing RNN, called EPNN, for nonconvex optimization are obtained, and two variants of the network by incorporating two augmented Lagrangian function techniques are proposed for seeking Karush-Kuhn-Tucker (KKT) points, especially local optima, of the problems. / Variational inequality (VI) can be viewed as a natural framework for unifying the treatment of equilibrium problems, and hence has applications across many disciplines. In addition, many typical problems are closely related to VI, including general VI (GVI), complementarity problem (CP), generalized CP (GCP) and optimization problem (OP). / Hu, Xiaolin. / "July 2007." / Adviser: Jun Wang. / Source: Dissertation Abstracts International, Volume: 69-02, Section: B, page: 1102. / Thesis (Ph.D.)--Chinese University of Hong Kong, 2007. / Includes bibliographical references (p. 193-207). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Electronic reproduction. [Ann Arbor, MI] : ProQuest Information and Learning, [200-] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Abstract in English and Chinese. / School code: 1307.

Li Wenjing. / "April 2002." / Thesis (Ph.D.)--Chinese University of Hong Kong, 2002. / Includes bibliographical references (p. 155-167). / Electronic reproduction. Hong Kong : Chinese University of Hong Kong, [2012] System requirements: Adobe Acrobat Reader. Available via World Wide Web. / Mode of access: World Wide Web. / Abstracts in English and Chinese.