Sara Porat

Learning automata from ordered examples

Connectionist learning models have had considerable empirical success, but it is hard to characterize exactly what they learn. The learning of finite-state languages (FSL) from example strings is a domain which has been extensively studied and might provide an opportunity to help understand connectionist learning. A major problem is that traditional FSL learning assumes the storage of all examples and thus violates connectionist principles. This paper presents a provably correct algorithm for inferring any minimum-state deterministic finite-state automata (FSA) from a complete ordered sample using limited total storage and without storing example strings. The algorithm is an iterative strategy that uses at each stage a current encoding of the data considered so far, and one single sample string. One of the crucial advantages of our algorithm is that the total amount of space used in the course of learning for encoding any finite prefix of the sample is polynomial in the size of the inferred minimum state deterministic FSA. The algorithm is also relatively efficient in time and has been implemented. More importantly, there is a connectionist version of the algorithm that preserves these properties. The connectionist version requires much more structure than the usual models and has been implemented using the Rochester Connectionist Simulator. We also show that no machine with finite working storage can iteratively identify the FSL from arbitrary presentations.

Learning Automata from Ordered Examples

Connectionist learning models have had considerable empirical success, but it is hard to characterize exactly what they learn. The learning of finite-state languages (FSL) from example strings is a domain which has been extensively studied and might provide an opportunity to help understand connectionist learning. A major problem is that traditional FSL learning assumes the storage of all examples and thus violates connectionist principles. This paper presents a provably correct algorithm for inferring any minimum-state deterministic finite-state automata (FSA) from a complete ordered sample using limited total storage and without storing example strings. The algorithm is an iterative strategy that uses at each stage a current encoding of the data considered so far, and one single sample string. One of the crucial advantages of our algorithm is that the total amount of space used in the course of learning for encoding any finite prefix of the sample is polynomial in the size of the inferred minimum state deterministic FSA. The algorithm is also relatively efficient in time and has been implemented. More importantly, there is a connectionist version of the algorithm that preserves these properties. The connectionist version requires much more structure than the usual models and has been implemented using the Rochester Connectionist Simulator. We also show that no machine with finite working storage can iteratively identify the FSL from arbitrary presentations.

The perception of others: inferring reputation from social media in the enterprise

The emergence of social media allows people to interact with others all over the world. During interaction, people leave many traces behind that can reveal things about themselves, or about how they perceive others: having many followers may indicate that one is an influencer; forum answers that gain high ranking, are likely to testify for expertise; people who gain high ranking in eCommerce sites are likely to be trustworthy. In this paper, we examine whether public online traces can be used for inferring the reputation of a person as perceived by others in relation to trustworthiness, influence, expertise, and impact. We describe a study performed on indicators of reputation that employees leave in a rich organizational social media platform. We compare different indicators, and report the results of an extensive user study with over 500 participants who provided their perception of thousands of others through a set of hypothetical scenarios.

Sealing, Encapsulation, and Mutability

Both encapsulation and immutability are important mechanisms, that support good software engineering practice. Encapsulation protects a variable against all kinds of access attempts from certain sections of the program. Immutability protects a variable only against write access attempts, irrespective of the program region from which these attempts are made. Taking mostly an empirical approach, we study these concepts and their interaction in JAVA. We propose code analysis techniques, which, using the new sealing information, can help to identify variables as encapsulated, immutable, or both.

Dynamic masking of application displays using OCR technologies

Industry coalitions are developing regulations to govern information sharing and to protect sensitive business data and the privacy of individuals. In many cases, these regulations make it impossible to outsource business operations, unless the companies have effective technologies to protect sensitive information. This paper addresses scenarios in which data servers and applications are owned and maintained on the premises of a company, and the service providers remotely access the data and the applications. We present a unique solution called Masking Gateway for Enterprises (MAGEN) that masks sensitive information appearing on application displays, without any interference with the applications that generate those screens. The major novelty lies in the utilization of optical character recognition (OCR) for analyzing and understanding application screens. Together with a comprehensive rule language, this approach makes it possible to characterize fields containing sensitive information and mask them according to predefined rules. The rule language is very flexible, abstract, and intuitive and is designed to cope with a vast set of policies and security needs. We describe the major challenges in implementing MAGEN and the results of experimenting with it in situations that occur in actual business settings. We outline techniques that optimize the OCR process to minimize latency and ensure robust operation.

Combined static and dynamic analysis for inferring program dependencies using a pattern language

One of the challenges when examining enterprise applications is the ability to understand the dependencies of these applications on external and internal resources such as database access or transaction activation. Inferring dependencies can be achieved using a static approach, a dynamic one or a combination of the two. Static analysis tools detect dependencies based on code investigation while dynamic tools detect dependencies based on runtime execution. The combination of these two approaches is essential for a complete and precise analysis. In this paper we present and illustrate a technique for inferring application dependencies on resources. The technique is based on a combined dynamic and static analysis. A pattern language is defined to enable the specification of dependencies as sequences of method invocations in the application code. Specifically, the sequences are patterns that constitute access to resources, e.g. databases, message queues, and control systems. We propose an algorithm for inferring application dependencies based on hybrid dynamic and static analysis that propagates information provided by dynamic analysis into the static analysis and back to the dynamic analysis. Empirical results from our implemented prototype are presented.