Visual object recognition is one of the key human capabilities that we would like machines to have. The problem is the following: given an image of an object (e.g. someone's face), predict its label (e.g. that person's name) from a set of possible object labels. The predominant approach to solving the recognition problem has been to learn a discriminative model, i.e. a model of the conditional probability $P(l|v)$ over possible object labels $l$ given an image $v$. Here we consider an alternative class of models, broadly referred to as \emph{generative models}, that learns the latent structure of the image so as to explain how it was generated. This is in contrast to discriminative models, which dedicate their parameters exclusively to representing the conditional distribution $P(l|v)$. Making finer distinctions among generative models, we consider a supervised generative model of the joint distribution $P(v,l)$ over image-label pairs, an unsupervised generative model of the distribution $P(v)$ over images alone, and an unsupervised \emph{reconstructive} model, which includes models such as autoencoders that can reconstruct a given image, but do not define a proper distribution over images. The goal of this thesis is to empirically demonstrate various ways of using these models for object recognition. Its main conclusion is that such models are not only useful for recognition, but can even outperform purely discriminative models on difficult recognition tasks. We explore four types of applications of generative/reconstructive models for recognition: 1) incorporating complex domain knowledge into the learning by inverting a synthesis model, 2) using the latent image representations of generative/reconstructive models for recognition, 3) optimizing a hybrid generative-discriminative loss function, and 4) creating additional synthetic data for training more accurate discriminative models. Taken together, the results for these applications support the idea that generative/reconstructive models and unsupervised learning have a key role to play in building object recognition systems.

In this thesis, we present a decision-theoretic framework for building decision support systems that incrementally elicit preferences of individual users over multiattribute outcomes and then provide recommendations based on the acquired preference information. By combining decision-theoretically sound modeling with effective computational techniques and certain user-centric considerations, we demonstrate the feasibility and potential of practical autonomous preference elicitation and recommendation systems. More concretely, we focus on decision scenarios in which a user can obtain any outcome from a finite set of available outcomes. The outcome is space is multiattribute; each outcome can be viewed as an instantiation of a set of attributes with finite domains. The user has preferences over outcomes that can be represented by a utility function. We assume that user preferences are generalized additively independent (GAI), and, therefore, can be represented by a GAI utility function. GAI utilities provide a flexible representation framework for structured preferences over multiattribute outcomes; they are less restrictive and, therefore, more widely applicable than additive utilities. In many decision scenarios with large and complex decision spaces (such as making travel plans or choosing an apartment to rent from thousands of available options), selecting the optimal decision can require a lot of time and effort on the part of the user. Since obtaining the user's complete utility function is generally infeasible, the decision support system has to support recommendation with partial preference information. We provide solutions for effective elicitation of GAI utilities in situations where a probabilistic prior about the user's utility function is available, and in situations where the system's uncertainty about user utilities is represented by maintaining a set of feasible user utilities. In the first case, we use Bayesian criteria for decision making and query selection. In the second case, recommendations (and query strategies) are based on the robust minimax regret criterion which recommends the outcome with the smallest maximum regret (with respect to all adversarial instantiations of feasible utility functions). Our proposed framework is implemented in the UTPref recommendation system that searches multiattribute product databases using the minimax regret criterion. UTPref is tested with a study involving 40 users interacting with the system. The study measures the effectiveness of regret-based elicitation, evaluates user comprehension and acceptance of minimax regret, and assesses the relative difficulty of different query types.

This thesis examines the issue of software evolution from a Requirements Engineering perspective. This perspective is founded on the premise that software evolution is best managed with reference to the requirements of a given software system. In particular, I follow the Requirements Problem approach to software development: the problem of developing software can be characterized as finding a specification that satisfies user requirements, subject to domain constraints. To enable this, I propose a shift from treating requirements as artifacts to treating requirements as design knowledge, embedded in knowledge bases. Most requirements today, when they exist in tangible form at all, are static objects. Such artifacts are quickly out of date and difficult to update. Instead, I propose that requirements be maintained in a knowledge base which supports knowledge-level operations for asserting new knowledge and updating existing knowledge. Consistency checks and entailment of new specifications is done automatically by answering simple queries. Maintaining a requirements knowledge base in parallel with running code means that changes precipitated by evolution are always addressed relative to the ultimate purpose of the system. This thesis begins with empirical studies which establish the nature of the requirements evolution problem. I use an extended case study of payment cards to motivate the following discussion. I begin at an abstract level, by introducing a requirements engineering knowledge base (REKB) using a functional specification. Since it is functional, the specifics of the implementation are left open. I then describe one implementation, using a reason-maintenance system, and show how this implementation can a) solve static requirements problems; b) help stakeholders bring requirements and implementation following a change in the requirements problem; c) propose paraconsistent reasoning to support inconsistency tolerance in the REKB. The end result of my work on the REKB is a tool and approach which can guide software developers and software maintainers in design and decision-making in the context of software evolution.

In this thesis, we present a decision-theoretic framework for building decision support systems that incrementally elicit preferences of individual users over multiattribute outcomes and then provide recommendations based on the acquired preference information. By combining decision-theoretically sound modeling with effective computational techniques and certain user-centric considerations, we demonstrate the feasibility and potential of practical autonomous preference elicitation and recommendation systems. More concretely, we focus on decision scenarios in which a user can obtain any outcome from a finite set of available outcomes. The outcome is space is multiattribute; each outcome can be viewed as an instantiation of a set of attributes with finite domains. The user has preferences over outcomes that can be represented by a utility function. We assume that user preferences are generalized additively independent (GAI), and, therefore, can be represented by a GAI utility function. GAI utilities provide a flexible representation framework for structured preferences over multiattribute outcomes; they are less restrictive and, therefore, more widely applicable than additive utilities. In many decision scenarios with large and complex decision spaces (such as making travel plans or choosing an apartment to rent from thousands of available options), selecting the optimal decision can require a lot of time and effort on the part of the user. Since obtaining the user's complete utility function is generally infeasible, the decision support system has to support recommendation with partial preference information. We provide solutions for effective elicitation of GAI utilities in situations where a probabilistic prior about the user's utility function is available, and in situations where the system's uncertainty about user utilities is represented by maintaining a set of feasible user utilities. In the first case, we use Bayesian criteria for decision making and query selection. In the second case, recommendations (and query strategies) are based on the robust minimax regret criterion which recommends the outcome with the smallest maximum regret (with respect to all adversarial instantiations of feasible utility functions). Our proposed framework is implemented in the UTPref recommendation system that searches multiattribute product databases using the minimax regret criterion. UTPref is tested with a study involving 40 users interacting with the system. The study measures the effectiveness of regret-based elicitation, evaluates user comprehension and acceptance of minimax regret, and assesses the relative difficulty of different query types.

This thesis examines the issue of software evolution from a Requirements Engineering perspective. This perspective is founded on the premise that software evolution is best managed with reference to the requirements of a given software system. In particular, I follow the Requirements Problem approach to software development: the problem of developing software can be characterized as finding a specification that satisfies user requirements, subject to domain constraints. To enable this, I propose a shift from treating requirements as artifacts to treating requirements as design knowledge, embedded in knowledge bases. Most requirements today, when they exist in tangible form at all, are static objects. Such artifacts are quickly out of date and difficult to update. Instead, I propose that requirements be maintained in a knowledge base which supports knowledge-level operations for asserting new knowledge and updating existing knowledge. Consistency checks and entailment of new specifications is done automatically by answering simple queries. Maintaining a requirements knowledge base in parallel with running code means that changes precipitated by evolution are always addressed relative to the ultimate purpose of the system. This thesis begins with empirical studies which establish the nature of the requirements evolution problem. I use an extended case study of payment cards to motivate the following discussion. I begin at an abstract level, by introducing a requirements engineering knowledge base (REKB) using a functional specification. Since it is functional, the specifics of the implementation are left open. I then describe one implementation, using a reason-maintenance system, and show how this implementation can a) solve static requirements problems; b) help stakeholders bring requirements and implementation following a change in the requirements problem; c) propose paraconsistent reasoning to support inconsistency tolerance in the REKB. The end result of my work on the REKB is a tool and approach which can guide software developers and software maintainers in design and decision-making in the context of software evolution.

Visual object recognition is one of the key human capabilities that we would like machines to have. The problem is the following: given an image of an object (e.g. someone's face), predict its label (e.g. that person's name) from a set of possible object labels. The predominant approach to solving the recognition problem has been to learn a discriminative model, i.e. a model of the conditional probability $P(l|v)$ over possible object labels $l$ given an image $v$. Here we consider an alternative class of models, broadly referred to as \emph{generative models}, that learns the latent structure of the image so as to explain how it was generated. This is in contrast to discriminative models, which dedicate their parameters exclusively to representing the conditional distribution $P(l|v)$. Making finer distinctions among generative models, we consider a supervised generative model of the joint distribution $P(v,l)$ over image-label pairs, an unsupervised generative model of the distribution $P(v)$ over images alone, and an unsupervised \emph{reconstructive} model, which includes models such as autoencoders that can reconstruct a given image, but do not define a proper distribution over images. The goal of this thesis is to empirically demonstrate various ways of using these models for object recognition. Its main conclusion is that such models are not only useful for recognition, but can even outperform purely discriminative models on difficult recognition tasks. We explore four types of applications of generative/reconstructive models for recognition: 1) incorporating complex domain knowledge into the learning by inverting a synthesis model, 2) using the latent image representations of generative/reconstructive models for recognition, 3) optimizing a hybrid generative-discriminative loss function, and 4) creating additional synthetic data for training more accurate discriminative models. Taken together, the results for these applications support the idea that generative/reconstructive models and unsupervised learning have a key role to play in building object recognition systems.

Image patches can be factorized into ‘shapelets’ that describe segmentation patterns, and palettes that describe how to paint the segments. This allows a flexible factorization of local shape (segmentation patterns) and appearance (palettes), which we argue is useful for tasks such as object and scene recognition. Here, we introduce the ‘shapelet’ model- a framework that is able to learn a library of ‘shapelet’ segmentation patterns to capture local shape, and hierarchical palettes of colors to capture appearance. Using a learned shapelet library, image patches can be analyzed using a variational technique to produce descriptors that separately describe local shape and local appearance. These descriptors can be used for high-level vision tasks, such as object and scene recognition. We show that the shapelet model is competitive with SIFT-based methods and structure element (stel) model variants on the object recognition datasets Caltech28 and Caltech101, and the scene recognition dataset All-I-Have-Seen.

Image patches can be factorized into ‘shapelets’ that describe segmentation patterns, and palettes that describe how to paint the segments. This allows a flexible factorization of local shape (segmentation patterns) and appearance (palettes), which we argue is useful for tasks such as object and scene recognition. Here, we introduce the ‘shapelet’ model- a framework that is able to learn a library of ‘shapelet’ segmentation patterns to capture local shape, and hierarchical palettes of colors to capture appearance. Using a learned shapelet library, image patches can be analyzed using a variational technique to produce descriptors that separately describe local shape and local appearance. These descriptors can be used for high-level vision tasks, such as object and scene recognition. We show that the shapelet model is competitive with SIFT-based methods and structure element (stel) model variants on the object recognition datasets Caltech28 and Caltech101, and the scene recognition dataset All-I-Have-Seen.

In dynamic domains, the state of the world may change in unexpected ways during the generation or execution of plans. Regardless of the cause of such changes, they raise the question of whether they interfere with ongoing planning efforts. Unexpected changes during plan generation may invalidate the current planning effort, while discrepancies between expected and actual state of the world during execution may render the executing plan invalid or sub-optimal, with respect to previously identified planning objectives. In this thesis we develop a general monitoring technique that can be used during both plan generation and plan execution to determine the relevance of unexpected changes and which supports recovery. This way, time intensive replanning from scratch in the new and unexpected state can often be avoided. The technique can be applied to a variety of objectives, including monitoring the optimality of plans, rather then just their validity. Intuitively, the technique operates in two steps: during planning the plan is annotated with additional information that is relevant to the achievement of the objective; then, when an unexpected change occurs, this information is used to determine the relevance of the discrepancy with respect to the objective. We substantiate the claim of broad applicability of this relevance-based technique by developing four concrete applications: generating optimal plans despite frequent, unexpected changes to the initial state of the world, monitoring plan optimality during execution, monitoring the execution of near-optimal policies in stochastic domains, and monitoring the generation and execution of plans with procedural hard constraints. In all cases, we use the formal notion of regression to identify what is relevant for achieving the objective. We prove the soundness of these concrete approaches and present empirical results demonstrating that in some contexts orders of magnitude speed-ups can be gained by our technique compared to replanning from scratch.

The goal of this research is to improve the breast cancer screening process based on magnetic resonance imaging (MRI). In a typical MRI breast examination, a radiologist is responsible for visually examining the MR images acquired during the examination and identifying suspect tissues for biopsy. It is known that if multiple radiologists independently analyze the same examinations and we biopsy any lesion that any of our radiologists flagged as suspicious then the overall screening process becomes more sensitive but less specific. Unfortunately cost factors prohibit the use of multiple radiologists for the screening of every breast MR examination. It is thought that instead of having a second expert human radiologist to examine each set of images, that the act of second reading of the examination can be performed by a computer-aided detection and diagnosis system. The research presented in this thesis is focused on the development of a computer-aided detection and diagnosis system for breast cancer screening from dynamic contrast-enhanced magnetic resonance imaging examinations. This thesis presents new computational techniques in supervised learning, unsupervised learning and classifier visualization. The techniques have been applied to breast MR lesion data and have been shown to outperform existing methods yielding a computer aided detection and diagnosis system with a sensitivity of 89% and a specificity of 70%.