Adaptive online gradient descent

We study the rates of growth of the regret in online convex optimization. First, we show that a simple extension of the algorithm of Hazan et al eliminates the need for a priori knowledge of the lower bound on the second derivatives of the observed functions. We then provide an algorithm, Adaptive Online Gradient Descent, which interpolates between the results of Zinkevich for linear functions and of Hazan et al for strongly convex functions, achieving intermediate rates between [square root T] and [log T]. Furthermore, we show strong optimality of the algorithm. Finally, we provide an extension of our results to general norms.

Adaptive online gradient descent

We study the rates of growth of the regret in online convex optimization. First, we show that a simple extension of the algorithm of Hazan et al eliminates the need for a priori knowledge of the lower bound on the second derivatives of the observed functions. We then provide an algorithm, Adaptive Online Gradient Descent, which interpolates between the results of Zinkevich for linear functions and of Hazan et al for strongly convex functions, achieving intermediate rates between [square root T] and [log T]. Furthermore, we show strong optimality of the algorithm. Finally, we provide an extension of our results to general norms.

Exponentiated gradient algorithms for conditional random fields and max-margin Markov Networks

Log-linear and maximum-margin models are two commonly-used methods in supervised machine learning, and are frequently used in structured prediction problems. Efficient learning of parameters in these models is therefore an important problem, and becomes a key factor when learning from very large data sets. This paper describes exponentiated gradient (EG) algorithms for training such models, where EG updates are applied to the convex dual of either the log-linear or max-margin objective function; the dual in both the log-linear and max-margin cases corresponds to minimizing a convex function with simplex constraints. We study both batch and online variants of the algorithm, and provide rates of convergence for both cases. In the max-margin case, O(1/ε) EG updates are required to reach a given accuracy ε in the dual; in contrast, for log-linear models only O(log(1/ε)) updates are required. For both the max-margin and log-linear cases, our bounds suggest that the online EG algorithm requires a factor of n less computation to reach a desired accuracy than the batch EG algorithm, where n is the number of training examples. Our experiments confirm that the online algorithms are much faster than the batch algorithms in practice. We describe how the EG updates factor in a convenient way for structured prediction problems, allowing the algorithms to be efficiently applied to problems such as sequence learning or natural language parsing. We perform extensive evaluation of the algorithms, comparing them to L-BFGS and stochastic gradient descent for log-linear models, and to SVM-Struct for max-margin models. The algorithms are applied to a multi-class problem as well as to a more complex large-scale parsing task. In all these settings, the EG algorithms presented here outperform the other methods.

Robust speech recognition using speech enhancement

Automatic Speech Recognition (ASR) has matured into a technology which is becoming more common in our everyday lives, and is emerging as a necessity to minimise driver distraction when operating in-car systems such as navigation and infotainment. In “noise-free” environments, word recognition performance of these systems has been shown to approach 100%, however this performance degrades rapidly as the level of background noise is increased. Speech enhancement is a popular method for making ASR systems more ro- bust. Single-channel spectral subtraction was originally designed to improve hu- man speech intelligibility and many attempts have been made to optimise this algorithm in terms of signal-based metrics such as maximised Signal-to-Noise Ratio (SNR) or minimised speech distortion. Such metrics are used to assess en- hancement performance for intelligibility not speech recognition, therefore mak- ing them sub-optimal ASR applications. This research investigates two methods for closely coupling subtractive-type enhancement algorithms with ASR: (a) a computationally-efficient Mel-filterbank noise subtraction technique based on likelihood-maximisation (LIMA), and (b) in- troducing phase spectrum information to enable spectral subtraction in the com- plex frequency domain. Likelihood-maximisation uses gradient-descent to optimise parameters of the enhancement algorithm to best fit the acoustic speech model given a word se- quence known a priori. Whilst this technique is shown to improve the ASR word accuracy performance, it is also identified to be particularly sensitive to non-noise mismatches between the training and testing data. Phase information has long been ignored in spectral subtraction as it is deemed to have little effect on human intelligibility. In this work it is shown that phase information is important in obtaining highly accurate estimates of clean speech magnitudes which are typically used in ASR feature extraction. Phase Estimation via Delay Projection is proposed based on the stationarity of sinusoidal signals, and demonstrates the potential to produce improvements in ASR word accuracy in a wide range of SNR. Throughout the dissertation, consideration is given to practical implemen- tation in vehicular environments which resulted in two novel contributions – a LIMA framework which takes advantage of the grounding procedure common to speech dialogue systems, and a resource-saving formulation of frequency-domain spectral subtraction for realisation in field-programmable gate array hardware. The techniques proposed in this dissertation were evaluated using the Aus- tralian English In-Car Speech Corpus which was collected as part of this work. This database is the first of its kind within Australia and captures real in-car speech of 50 native Australian speakers in seven driving conditions common to Australian environments.

Robot navigation in sensor space

This thesis investigates the problem of robot navigation using only landmark bearings. The proposed system allows a robot to move to a ground target location specified by the sensor values observed at this ground target posi- tion. The control actions are computed based on the difference between the current landmark bearings and the target landmark bearings. No Cartesian coordinates with respect to the ground are computed by the control system. The robot navigates using solely information from the bearing sensor space. Most existing robot navigation systems require a ground frame (2D Cartesian coordinate system) in order to navigate from a ground point A to a ground point B. The commonly used sensors such as laser range scanner, sonar, infrared, and vision do not directly provide the 2D ground coordi- nates of the robot. The existing systems use the sensor measurements to localise the robot with respect to a map, a set of 2D coordinates of the objects of interest. It is more natural to navigate between the points in the sensor space corresponding to A and B without requiring the Cartesian map and the localisation process. Research on animals has revealed how insects are able to exploit very limited computational and memory resources to successfully navigate to a desired destination without computing Cartesian positions. For example, a honeybee balances the left and right optical flows to navigate in a nar- row corridor. Unlike many other ants, Cataglyphis bicolor does not secrete pheromone trails in order to find its way home but instead uses the sun as a compass to keep track of its home direction vector. The home vector can be inaccurate, so the ant also uses landmark recognition. More precisely, it takes snapshots and compass headings of some landmarks. To return home, the ant tries to line up the landmarks exactly as they were before it started wandering. This thesis introduces a navigation method based on reflex actions in sensor space. The sensor vector is made of the bearings of some landmarks, and the reflex action is a gradient descent with respect to the distance in sensor space between the current sensor vector and the target sensor vec- tor. Our theoretical analysis shows that except for some fully characterized pathological cases, any point is reachable from any other point by reflex action in the bearing sensor space provided the environment contains three landmarks and is free of obstacles. The trajectories of a robot using reflex navigation, like other image- based visual control strategies, do not correspond necessarily to the shortest paths on the ground, because the sensor error is minimized, not the moving distance on the ground. However, we show that the use of a sequence of waypoints in sensor space can address this problem. In order to identify relevant waypoints, we train a Self Organising Map (SOM) from a set of observations uniformly distributed with respect to the ground. This SOM provides a sense of location to the robot, and allows a form of path planning in sensor space. The navigation proposed system is analysed theoretically, and evaluated both in simulation and with experiments on a real robot.

Decision support system for contractor pre-qualification : artificial neural network model

The selection criteria for contractor pre-qualification are characterized by the co-existence of both quantitative and qualitative data. The qualitative data is non-linear, uncertain and imprecise. An ideal decision support system for contractor pre-qualification should have the ability of handling both quantitative and qualitative data, and of mapping the complicated nonlinear relationship of the selection criteria, such that rational and consistent decisions can be made. In this research paper, an artificial neural network model was developed to assist public clients identifying suitable contractors for tendering. The pre-qualification criteria (variables) were identified for the model. One hundred and twelve real pre-qualification cases were collected from civil engineering projects in Hong Kong, and eighty-eight hypothetical pre-qualification cases were also generated according to the “If-then” rules used by professionals in the pre-qualification process. The results of the analysis totally comply with current practice (public developers in Hong Kong). Each pre-qualification case consisted of input ratings for candidate contractors’ attributes and their corresponding pre-qualification decisions. The training of the neural network model was accomplished by using the developed program, in which a conjugate gradient descent algorithm was incorporated for improving the learning performance of the network. Cross-validation was applied to estimate the generalization errors based on the “re-sampling” of training pairs. The case studies show that the artificial neural network model is suitable for mapping the complicated nonlinear relationship between contractors’ attributes and their corresponding pre-qualification (disqualification) decisions. The artificial neural network model can be concluded as an ideal alternative for performing the contractor pre-qualification task.

Optimising figure of merit for phonetic spoken term detection

This paper introduces a novel technique to directly optimise the Figure of Merit (FOM) for phonetic spoken term detection. The FOM is a popular measure of sTD accuracy, making it an ideal candiate for use as an objective function. A simple linear model is introduced to transform the phone log-posterior probabilities output by a phe classifier to produce enhanced log-posterior features that are more suitable for the STD task. Direct optimisation of the FOM is then performed by training the parameters of this model using a non-linear gradient descent algorithm. Substantial FOM improvements of 11% relative are achieved on held-out evaluation data, demonstrating the generalisability of the approach.

Modeling of a magneto-rheological (MR) fluid damper using a self tuning fuzzy mechanism

A magneto-rheological (MR) fluid damper is a semi-active control device that has recently begun to receive more attention in the vibration control community. However, the inherent nonlinear nature of the MR fluid damper makes it challenging to use this device to achieve high damping control system performance. Therefore the development of an accurate modeling method for a MR fluid damper is necessary to take advantage of its unique characteristics. Our goal was to develop an alternative method for modeling a MR fluid damper by using a self tuning fuzzy (STF) method based on neural technique. The behavior of the researched damper is directly estimated through a fuzzy mapping system. In order to improve the accuracy of the STF model, a back propagation and a gradient descent method are used to train online the fuzzy parameters to minimize the model error function. A series of simulations had been done to validate the effectiveness of the suggested modeling method when compared with the data measured from experiments on a test rig with a researched MR fluid damper. Finally, modeling results show that the proposed STF interference system trained online by using neural technique could describe well the behavior of the MR fluid damper without need of calculation time for generating the model parameters.

Fourier Lucas-Kanade Algorithm

In this paper we propose a framework for both gradient descent image and object alignment in the Fourier domain. Our method centers upon the classical Lucas & Kanade (LK) algorithm where we represent the source and template/model in the complex 2D Fourier domain rather than in the spatial 2D domain. We refer to our approach as the Fourier LK (FLK) algorithm. The FLK formulation is advantageous when one pre-processes the source image and template/model with a bank of filters (e.g. oriented edges, Gabor, etc.) as: (i) it can handle substantial illumination variations, (ii) the inefficient pre-processing filter bank step can be subsumed within the FLK algorithm as a sparse diagonal weighting matrix, (iii) unlike traditional LK the computational cost is invariant to the number of filters and as a result far more efficient, and (iv) this approach can be extended to the inverse compositional form of the LK algorithm where nearly all steps (including Fourier transform and filter bank pre-processing) can be pre-computed leading to an extremely efficient and robust approach to gradient descent image matching. Further, these computational savings translate to non-rigid object alignment tasks that are considered extensions of the LK algorithm such as those found in Active Appearance Models (AAMs).

Utilizing voting systems for ranking user tweets

Twitter is a very popular social network website that allows users to publish short posts called tweets. Users in Twitter can follow other users, called followees. A user can see the posts of his followees on his Twitter profile home page. An information overload problem arose, with the increase of the number of followees, related to the number of tweets available in the user page. Twitter, similar to other social network websites, attempts to elevate the tweets the user is expected to be interested in to increase overall user engagement. However, Twitter still uses the chronological order to rank the tweets. The tweets ranking problem was addressed in many current researches. A sub-problem of this problem is to rank the tweets for a single followee. In this paper we represent the tweets using several features and then we propose to use a weighted version of the famous voting system Borda-Count (BC) to combine several ranked lists into one. A gradient descent method and collaborative filtering method are employed to learn the optimal weights. We also employ the Baldwin voting system for blending features (or predictors). Finally we use the greedy feature selection algorithm to select the best combination of features to ensure the best results.

Exponentiated gradient algorithms for large-margin structured classification

We consider the problem of structured classification, where the task is to predict a label y from an input x, and y has meaningful internal structure. Our framework includes supervised training of Markov random fields and weighted context-free grammars as special cases. We describe an algorithm that solves the large-margin optimization problem defined in [12], using an exponential-family (Gibbs distribution) representation of structured objects. The algorithm is efficient—even in cases where the number of labels y is exponential in size—provided that certain expectations under Gibbs distributions can be calculated efficiently. The method for structured labels relies on a more general result, specifically the application of exponentiated gradient updates [7, 8] to quadratic programs.

Infinite-horizon policy-gradient estimation

Gradient-based approaches to direct policy search in reinforcement learning have received much recent attention as a means to solve problems of partial observability and to avoid some of the problems associated with policy degradation in value-function methods. In this paper we introduce GPOMDP, a simulation-based algorithm for generating a biased estimate of the gradient of the average reward in Partially Observable Markov Decision Processes (POMDPs) controlled by parameterized stochastic policies. A similar algorithm was proposed by Kimura, Yamamura, and Kobayashi (1995). The algorithm's chief advantages are that it requires storage of only twice the number of policy parameters, uses one free parameter β ∈ [0,1) (which has a natural interpretation in terms of bias-variance trade-off), and requires no knowledge of the underlying state. We prove convergence of GPOMDP, and show how the correct choice of the parameter β is related to the mixing time of the controlled POMDP. We briefly describe extensions of GPOMDP to controlled Markov chains, continuous state, observation and control spaces, multiple-agents, higher-order derivatives, and a version for training stochastic policies with internal states. In a companion paper (Baxter, Bartlett, & Weaver, 2001) we show how the gradient estimates generated by GPOMDP can be used in both a traditional stochastic gradient algorithm and a conjugate-gradient procedure to find local optima of the average reward. ©2001 AI Access Foundation and Morgan Kaufmann Publishers. All rights reserved.

Selection procedures for module discovery : exploring evolutionary algorithms for cognitive science

Evolutionary algorithms are playing an increasingly important role as search methods in cognitive science domains. In this study, methodological issues in the use of evolutionary algorithms were investigated via simulations in which procedures were systematically varied to modify the selection pressures on populations of evolving agents. Traditional roulette wheel, tournament, and variations of these selection algorithms were compared on the “needle-in-a-haystack” problem developed by Hinton and Nowlan in their 1987 study of the Baldwin effect. The task is an important one for cognitive science, as it demonstrates the power of learning as a local search technique in smoothing a fitness landscape that lacks gradient information. One aspect that has continued to foster interest in the problem is the observation of residual learning ability in simulated populations even after long periods of time. Effective evolutionary algorithms balance their search effort between broad exploration of the search space and in-depth exploitation of promising solutions already found. Issues discussed include the differential effects of rank and proportional selection, the tradeoff between migration of populations towards good solutions and maintenance of diversity, and the development of measures that illustrate how each selection algorithm affects the search process over generations. We show that both roulette wheel and tournament algorithms can be modified to appropriately balance search between exploration and exploitation, and effectively eliminate residual learning in this problem.