Abstract: In most papers establishing consistency for learning algorithms it is assumed that the observations used for training are realizations of an i.i.d. process. In this paper we go far beyond this classical framework by showing that support vector machines (SVMs) essentially only require that the data-generating process satisfies a certain law of large numbers. We then consider the learnability of SVMs for $\a$-mixing (not necessarily stationary) processes for both classification and regression, where for the latter we explicitly allow unbounded noise.

Abstract: Purpose – As companies extend supply chains via direct delivery to consumers, supply chain efficiency depends upon the usability of the online ordering system. The purpose of this paper is to focus on customer order cycle efficiency gains through the “learnability” of web sites.Design/methodology/approach – The paper analyzes empirical data using nonlinear regression from seven firms and over 4,000 customers to examine how order time – an important performance metric – changes within an online grocery ordering environment.Findings – The evidence supports various forms of power‐law learning for web‐based ordering (i.e. the first few orders involve substantial learning). However, significant differences exist between web sites, and a portion of the ordering time may be irreducible.Research limitations/implications – The research lends insight into how web sites influence last‐mile supply chain efficiency via differing learning rates in the order cycle. Perceptual measures were used in order to assess custom...

Abstract: Acknowledgements I consider myself very lucky to have had the support when writing this thesis, whether directly or indirectly, from the following people: Essential to starting and finishing this work is Paul Boersma. I thank him for his support as much as for challenging me, and for enabling me to push the limits. I thank my reading committee Haike Jacobs, René Kager, Wolfgang Kehrein, Norval Smith, and Paul Smolensky for their valuable comments; especially Wolfgang and Norval for enduring tedious questions about phonology. My paranymphs Maren Pannemann and Petra Jongmans never failed in supporting me; they prevented me from going nuts in quite some moments of panic. I wouldn't have known what to do without them, and my Dutch would be much worse if it weren't for them. Furthermore, I was very happy to be part of the phonological reading groups in Utrecht and Leiden. It was always fun and helpful to discuss anything phono with and all the other people who attended. Being part of the Sound Circle organizing committee with Ivana Brasileiro, Ania Kijak, and Shakuntala Mahanta was, and still is, great. The core of the IFA as I got to know it made me always feel welcome. Ton Wempe en Dirk Jaasma took more than good care of me, not to forget my fellow gratiën Irene Jacobi and Wieneke Wesseling. Being a phonologist among phoneticians is not so bad at all. In the bigger picture, I thank everyone at the ACLC for contributing to a stimulating work environment, and all AiOs for being an enjoyable group open for discussion. Being responsible for my beginning interest in science, I like to thank Janet Grijzenhout and Martin Krämer, who have always been there with good advise for me beyond Düsseldorf. There are many non-phonologist linguists I need to thank for a lot inside and outside linguistics: Martin Salzmann for having the first fun-nights out in Amsterdam; Berit Gehrke for many more fun-nights out and in; Mika Poss for being there in Boston; Silke Hamann for being a role model, and Robert Cloutier for his patience and help with my English. My fellow mappreviators need to be mentioned, too: Roland Pfau for being my other role model; Rachel Selbach and Karina Hof for brightening my day with their eloquence and quick cheer-ups; and Rafael Fischer for being a fun office mate and utterly cheek. Completely outside of linguistics, there are …

Abstract: This paper presents initial research on a new learnability assessment methodology. We propose the use of electroencephalography (EEG) to further improve usability testing. We discovered whether and to what extend there is a correlation between brainwave patterns and the learnability of the software used. Our central hypothesis is that learnability can be assessed by analyzing the rise and fall of specific frequency bands in electroencephalographic recordings. In order to collect empirical evidence for our hypothesis, we conducted an experiment with N=32 participants. We developed a test environment comprising a low-cost EEG system and developed software for analysis and testing. Based on our findings, we consider our EEG-based learnability test applicable, either as a pre-test - in order to determine whether further testing is necessary - or as an augmenting method during standard usability testing. The users' emotions, registered on the EEG, can be applied as a baseline for detecting possible usability difficulties and employed in the development of a biological rapid-usability method for accessibility assessment.

Abstract: In this paper, we consider Kearns' [4] Statistical Query Model of learning. It is well known [3] that the number of statistical queries, needed for "weakly learning" an unknown target concept (i.e. for gaining significant advantage over random guessing) is polynomially related to the so-called Statistical Query dimension of the concept class. In this paper, we provide a similar characterization for "strong learning" where the learners final hypothesis is required to approximate the unknown target concept up to a small rate of misclassification. The quantity that characterizes strong learnability in the Statistical Query model is a surprisingly close relative of (though not identical to) the Statistical Query dimension. For the purpose of proving the main result, we provide other characterizations of strong learnability which are given in terms of covering numbers and related notions. These results might find some interest in their own right. All characterizations are purely information-theoretical and ignore computational issues.

Abstract: Currently, equivalence partitioning and boundary value analysis are taught at La Trobe University using Myers' original representation of these black-box testing methods. We previously proposed an alternative representation called atomic rules. In this paper we present the statistical results of two similar experiments that examine which of these approaches enable students to write more complete and correct black-box test sets and which approach students prefer to use. We compare the results of these experiments and discuss how the results could change the teaching of black-box testing methods at La Trobe University and in industry.

Abstract: In this article I provide evidence that despite frequently cited differences between child first language (L1) and adult second language (L2) speakers in overt behavior (performance) during grammat...

Abstract: A model for learning in the limit is defined where a (so-called iterative) learner gets all positive examples from the target language, tests every new conjecture with a teacher (oracle) if it is a subset of the target language (and if it is not, then it receives a negative counterexample), and uses only limited long-term memory (incorporated in conjectures). Three variants of this model are compared: when a learner receives least negative counterexamples, the ones whose size is bounded by the maximum size of input seen so far, and arbitrary ones. A surprising result is that sometimes absence of bounded counterexamples can help an iterative learner whereas arbitrary counterexamples are useless. We also compare our learnability model with other relevant models of learnability in the limit, study how our model works for indexed classes of recursive languages, and show that learners in our model can work in non-U-shaped way-never abandoning the first right conjecture.

Abstract: The verb inflections of Bengali underwent a series of phonological change between 10th and 18th centuries, which gave rise to several modern dialects of the language. In this paper, we offer a functional explanation for this change by quantifying the functional pressures of ease of articulation, perceptual contrast and learnability through objective functions or constraints, or both. The multi-objective and multi-constraint optimization problem has been solved through genetic algorithm, whereby we have observed the emergence of Pareto-optimal dialects in the system that closely resemble some of the real ones.

Abstract: This paper describes an approach to support non-visual exploration of graphically represented information. We used a hierarchical structure to organize the information encoded in a relational diagram and designed two alternative audio-only interfaces for presenting the hierarchy, each employing different levels of verbosity. We report on an experimental study that assessed the viability of our proposed approach as well as the efficiency and learnability of each interface. Our results show that the relational information encoded in a diagram could be non-visually navigated and explored through a hierarchy, and that substituting verbal descriptions of parts of such information with nonverbal sounds significantly improve performance without compromising comprehension.

Abstract: No systematic empirical study on investigating the effects of performing task variants on user cognitive strategy and behaviour in usability tests and on learnability of the system being tested has been documented in the literature. The current use-inspired basic research work aims to identify the underlying cognitive mechanisms and the practical implications of this specific endeavour. The focus of our work was to assess user rationality and system learnability. The software application tested was a multilingual learning resource repository. Eleven German and eleven Slovenian participants were involved in two user tests (UTs). Usability problems (UPs) identified in two quasi-isomorphic tasks were categorized with respect to a scheme of associated skills. Actions of the two tasks of each of the 22 users were segmented and coded according to a scheme of cognitive activities. Results showed that generally the users adopted different strategies for working out the given task and its variant, and that the system could be proved learnable. User Rational Action Model and implications for future research on user tests are inferred.

Abstract: We present four computational experiments that investigate the impact of null subjects (pro-drop) on the learnability of languages with a basic Subject-Verb-Object (SVO) word order and varying amounts of morphological marking on their nouns and verbs. The simulations show that the effect of pro-drop on language learnability is limited as long as some morphological marking is present. Contrary to expectation, rich agreement markers are no more useful in the simulations than nominal case markers or verbal Tense/Aspect/Modality markers. In the absence of morphological marking, however, pro-drop leads to severe learnability problems in the simulations: overall performance on this language type is significantly worse (Experiment 1); additional exposure to language data is not as useful as with other types (Experiment 2); novel words are more problematic in this type (Experiment 3); and noun/verb homonyms also decrease performance for this type (Experiment 4). An analysis of the simulations shows that the main problem is accurately distinguishing nouns from verbs. These results suggests that the combination of pro-drop and no morphological marking should be unattested among natural languages. To test this hypothesis we first survey various creole languages as they are SVO and typically lack morphological markers. However, cross-linguistic data shows that creole languages do not allow pro-drop unless they have also developed agreement markers. We then discuss Mandarin Chinese because it allows widespread pro-drop and features only minimal morphological marking. A closer look at the language reveals that Mandarin provides quite reliable cues for identifying nouns and verbs in the language. Crucially, these cues are acquired very early by children learning Mandarin. Similarly, children only very rarely use nouns as verbs (or vice versa)---unlike in English where pro-drop is not possible. Two other unusual properties of Mandarin Chinese that are also compatible with our experimental results are the relatively early acquisition of verbs and the presence of relatively frequent noun/verb homonymy. Mandarin is thus not a counter-example to the results of the simulations. We end by situating our work in relation to various other approaches, such as the Competition Model, Optimality Theory, and probabilistic linguistics.

Abstract: In the study of acquisition and learnability in Optimality Theory (OT; Prince and Smolensky 1993/2004) learning is characterized in terms of changes in constraint ranking. Learners begin with a ranking of Markedness constraints above Faithfulness constraints, and rerank them on the basis of evidence from the target language. A theory of learnability that accounts for the human acquisition process should both converge on the correct final grammar, and model the path that learners take to get there. The Gradual Learning Algorithm (GLA; Boersma 1998, Boersma and Hayes 2001) can model some, but not all aspects of the learning path, and as we will show, is non-convergent. The Constraint Demotion Algorithm (Tesar and Smolensky 1998) is convergent, but non-gradual. In this paper we argue that the search for a gradual convergent learner may be aided by replacing Optimality Theory’s constraint ranking with numerical weighting, returning in this respect to OT's predecessor Harmonic Grammar (HG; Legendre et al. 1990, Smolensky and Legendre 2006). We demonstrate the advantages of weighting by using a minimally modified version of the GLA implemented by Boersma and Weenink (2006) that learns Harmonic Grammars, rather than OT grammars. In the next section, we provide examples of two types of gradualness in the first language acquisition of phonology, before moving on to the question of how this gradualness can be modeled in learnability theory.

Abstract: It is commonly assumed that non-native optionality, i.e. where two competing grammars exist in the mental representation of L2 learners, is a common feature of developing grammars even at advanced proficiency levels (White 1991, 1992, Eubank 1994, Sorace, 1993, 1999, 2000, Prevost and White 2000). In this study we focus on the source of such optionality arguing that certain errors found in advanced non-native grammars cannot be sufficiently accounted for as simple transfers from the learner’s L1. Our study builds upon the observation that real optionality in native grammars is in fact difficult to verify (Papp 2000, Parodi and Tsimpli 2005) and that whereas two possible variations of the structure may coexist in the target grammar, the contexts in which the forms are used are not easily identified. Consequently, the linguistic evidence from which L2 learners create grammatical assumptions can be quite ambiguous. Given such obvious lack of robustness in the input, the learnability task is made considerably more difficult and presumably learners will face longer periods of grammatical indeterminacy even at advanced levels of proficiency. Even though it is commonly assumed that optional constructions surface in both incomplete and divergent L2 near-native grammars, the source of such optionality remains unclear, and it is generally assumed that learners may revert to their native language when they find difficulty in inducing the rules of the target grammar (Sorace 1993, Papp 2000). In this respect the acquisition of focus in Spanish is a good test ground because the native input is ambiguous, there is more than one way of marking focus (Zubizarreta 1998, Dominguez 2004), and also because advanced learners encounter problems acquiring the pragmatic conditions that constrain word order alterations in focused sentences (Ocampo 1990, Hertel 2003, De Miguel 1993, Lozano 2006).

Abstract: This paper investigates the learnability by positive examples in the sense of Gold of Pregroup Grammars In a first part, Pregroup Grammars are presented and a new parsing strategy is proposed Then, theoretical learnability and non-learnability results for subclasses of Pregroup Grammars are proved In the last two parts, we focus on learning Pregroup Grammars from a special kind of input called feature-tagged examples A learning algorithm based on the parsing strategy presented in the first part is given Its validity is proved and its properties are examplified

Abstract: FOSSILIZED SECOND LANGUAGE GRAMMARS Florencia Franceschina Amsterdam: Benjamins, 2005 Pp xxiv + 288 $13800 clothThis monograph by Franceschina presents an empirical study of what is known as Orwell's problem: a learnability problem manifested as lack of learning in spite of exposure to abundant and unambiguous positive input evidence The study is premised on three main assumptions: (a) Universal Grammar (UG) operates normally in all types of natural language acquisition, (b) first language (L1)-selected parameterized functional features (PFFs) are available in the acquisition of languages other than the L1, and (c) there is a critical period for the acquisition of PFFs Following Hawkins and Chan's (1997) failed functional features hypothesis (FFFH), Franceschina hypothesizes that in adult SLA, nativelike knowledge of PFFs will be restricted to the subset instantiated in the L1 Using gender marking in second language (L2) Spanish as a test case, the study addresses the following questions: (a) Can adult learners acquire nativelike knowledge of grammatical gender in the L2? (b) In adult L2 learners, is the possibility of nativelike attainment in the area of grammatical gender determined by the learner's L1? (c) What might prevent near-natives from reaching the same endstate knowledge as L1 speakers? Two experimental groups of adult near-natives with contrasting L1s (+gen vs −gen) relative to the target language and a control group of native speakers were formed and subsequently subjected to six experimental tasks that tapped into comprehension, production, and metalinguistic judgments The data from each test task separately underwent quantitative analyses, and the results consistently pointed to an advantage for the +gen group and, conversely, a disadvantage for the −gen group, thereby confirming the FFFH and, hence, Franceschina's own hypothesis

Abstract: Nowadays, subject testing represents a well-established methodology to evaluate different properties of driver information systems and driver assistance systems. Among several criteria, learnability is one important system property. User and usage strategies are dependent on the subject’s learning state, for example, to switch attendance between driving task and operation of a driver information system. Therefore it is wishful that the user acquires a model of the system, for example learns as quickly as possible. Also, the intended usage of driver assistance systems in given driving situations is influenced by the user’s experience. A suitable way to investigate related questions is to conduct a typical learning experiment and to analyse data with the given methodology. In this type of experiment, the familiarity and training state of the subject are set as independent variable. Beside learnablity, other properties of human-machine interaction are to be investigated and evaluated. In this case, however, the learning effectuated by a subject is an important dependent variable or even noise in sense of measurement theory that might cover a given main effect. After some empirical examples, possible solutions will be discussed that help to manage this problem with justifiable expense.

Abstract: In language learning, strong relationships between Gold-style models and query models have recently been observed: in some quite general setting Gold-style learners can be replaced by query learners and vice versa, without loss of learning capabilities. These 'equalities' hold in the context of learning indexable classes of recursive languages. Former studies on Gold-style learning of such indexable classes have shown that, in many settings, the enumerability of the target class and the recursiveness of its languages are crucial for learnability. Moreover, studying query learning, non-indexable classes have been mainly neglected up to now. So it is conceivable that the recently observed relations between Gold-style and query learning are not due to common structures in the learning processes in both models, but rather to the enumerability of the target classes or the recursiveness of their languages. In this paper, the analysis is lifted onto the context of learning arbitrary classes of recursively enumerable languages. Still, strong relationships between the approaches of Gold-style and query learning are proven, but there are significant changes to the former results. Though in many cases learners of one type can still be replaced by learners of the other type, in general this does not remain valid vice versa. All results hold even for learning classes of recursive languages, which indicates that the recursiveness of the languages is not crucial for the former 'equality' results. Thus we analyze how constraints on the algorithmic structure of the target class affect the relations between two approaches to language learning.

Abstract: In this paper we investigate the problem of classifying vertices of a directed graph according to an unknown directed cut. We first consider the usual setting in which the directed cut is fixed. However, even in this setting learning is not possible without in the worst case needing the labels for the whole vertex set. By considering the size of the minimum path cover as a fixed parameter, we derive positive learnability results with tight performance guarantees for active, online, as well as PAC learning. The advantage of this parameter over possible alternatives is that it allows for an a priori estimation of the total cost of labelling all vertices. The main result of this paper is the analysis of learning directed cuts that depend on a hidden and changing context.

Abstract: In this paper we investigate the problem of classifying vertices of a directed graph according to an unknown directed cut. We first consider the usual setting in which the directed cut is fixed. However, even in this setting learning is not possible without in the worst case needing the labels for the whole vertex set. By considering the size of the minimum path cover as a fixed parameter, we derive positive learnability results with tight performance guarantees for active, online, as well as PAC learning. The advantage of this parameter over possible alternatives is that it allows for an a priori estimation of the total cost of labelling all vertices. The main result of this paper is the analysis of learning directed cuts that depend on a hidden and changing context.

Abstract: As some cognitive research suggests, in the process of learning languages, in addition to overtexplicit negative evidence, a child often receives covertexplicit evidence in form of corrected or rephrased sentences. In this paper, we suggest one approach to formalization of overt and covert evidence within the framework of one-shotlearners via subset and membership queries to a teacher (oracle). We compare and explore general capabilities of our models, as well as complexity advantages of learnability models of one type over models of other types, where complexity is measured in terms of number of queries. In particular, we establish that "correcting" positive examples give sometimes more power to a learner than just negative (counter)examples and access to full positive data.

Abstract: Error bounds for decision trees are generally based on depth or breadth of the tree In this paper, we propose a bound for error rate that depends both on the depth and the breadth of a specific decision tree constructed from the training samples This bound is derived from sample complexity estimate based on PAC learnability The proposed bound is compared with other traditional error bounds on several machine learning benchmark data sets as well as on an image data set used in Content Based Image Retrieval (CBIR) Experimental results demonstrate that the proposed bound gives tighter estimation of the empirical error