No description available.

When old/new recognition memory is tested with equal numbers of studied and non-studied items and no rewards or instructions that favour one response over the other, there is no obvious reason for response bias. In line with this, Canadian undergraduates have shown, on average, a neutral response bias when we tested them on recognition of common English words. By contrast, most subjects we have tested on recognition of richly detailed images have shown a conservative bias: they more often erred by missing a studied image than by judging a non-studied image as studied. Here, in an effort to better understand these materials-based bias effects (MBBEs), we examined changes in hit and false alarm (FA) rates (and in sensitivity and bias) from the first to fourth quartile of a recognition memory test in eight experiments in which undergraduates studied words and/or images of paintings. Response bias for images tended to increase across quartiles, whereas bias for words showed no consistent pattern across quartiles. This pattern could be described as an increase in the MBBE over the course of the test, but the underlying patterns for hits and FAs are not easily reconciled with this interpretation. Hit rates decreased over the course of the test for both materials types, with that decline tending to be steeper for images than words. For words, FA rates tended to increase across quartiles, whereas for paintings FA rates did not increase across quartiles. We discuss implications of these findings for theoretical accounts of the MBBE. / Graduate

Speech recognition is now a key topic in computer science with the proliferation of voice-activated assistants, and voice-enabled devices. Many companies over a speech recognition service for developers to use to enable smart devices and services. These speech-to-text systems, however, have significant room for improvement, especially in domain specific speech. IBM's Watson speech-to-text service attempts to support domain specific uses by allowing users to upload their own training data for making custom models that augment Watson's general model. This requires deciding a strategy for picking the training model. This thesis experiments with different training choices for custom language models that augment Watson's speech to text service. The results show that using recent utterances is the best choice of training data in our use case of Digital Democracy. We are able to improve speech recognition accuracy by 2.3% percent over the control with no custom model. However, choosing training utterances most specific to the use case is better when large enough volumes of such training data is available.

The purpose of this dissertation is to investigate the feasibility of using neural networks in conjunction with the Fourier Modified Direct Mellin Transform (FMDMT) for the recognition of ship targets. The FMDMT is a modification of the Direct Mellin Transform for digital implementations, and is applied to the magnitudes of the Discrete Fourier Transforms (DFT) of range profiles of ships. Necessity for the use of the FMDMT is corroborated by the fact that features can be extracted from the range profiles of targets, regardless of target aspect angle. Variation in aspect angle results in variation of the independent variable. Feature extraction is made possible by the scale invariant properties of the Mellin Transform. Substantial emphasis was placed on preprocessing techniques applied in the implementation of the FMDMT on simulated range profiles and in particular, real ship profiles. The FMDMT was thus examined extensively and utilised as it was developed and demonstrated in [20]. At the completion of this examination, the recognition procedures and methods were applied on simulated data with the aid of a radar simulator developed and adapted for this dissertation. Results of the recognition of simulated ship targets were scrutinized closely and recorded. Employment of this procedure afforded the ability to compare the recognition results for real ship data with those of simulated ship data at a later stage. Acquisition of a large database of ship profiles was made successful by a ship target data capture plan implemented at the Institute for Maritime Technology (IMT) in Simon's Town. The database included the radar range profile data for the SAS Protea and the Outeniqua, which carried out several successful full circular manoeuvres in the line of sight of the search radar utilised (Raytheon). The relevant ships performed these circular manoeuvres in order that the acquired data incorporate radar range profiles of the relevant ships at most aspect angles from 0 degrees to 360 degrees. Extensive and thorough testing of the performance of the FMDMT would thus be possible since every possible aspect angle would be scrutinized. Preprocessing of data and recognition of targets was implemented in exactly the same manner and order as was the case with the simulated ship data. Extensive examination of the FMDMT revealed that the MDMT should only be applied to one side of a real and even Fourier Transform of a ship target. Literature on the FMDMT had failed to elaborate on this point. Comparison of the recognition results for real and simulated data, indicates a great similarity in success, thus validating the methods and procedures described theoretically and adopted practically for preprocessing of the radar range profiles and recognition of the targets. In order to demonstrate the feasibility of ship target recognition using the procedures and methods incorporated in the dissertation, real ship data for an entire range of different ships should be acquired in the same manner as indicated above. Bibliography: pages 117-118.

Bibliography: pages 239-248. / This thesis addresses a practical problem. The problem concerns the evaluation of 'identification parades', or 'lineups', which are frequently used by police to secure evidence of identification. It is well recognised that this evidence is frequently unreliable, and has led on occasion to tragic miscarriages of justice. A review of South African law is conducted and reported in the thesis, and shows that the legal treatment of identification parades centres on the requirement that parades should be composed of people of similar appearance to the suspect. I argue that it is not possible, in practice, to assess whether this requirement has been met and that this is a significant failing. Psychological work on identification parades includes the development of measures of parade fairness, and the investigation of alternate lineup structures. Measures of parade fairness suggested in the literature are indirectly derived, though; and I argue that they fail to address the question of physical similarity. In addition, I develop ways of reasoning inferentially (statistically) with measures of parade fairness, and suggest a new measure of parade fairness. The absence of a direct measure of similarity constitutes the rationale for the empirical component of the thesis. I propose a measure of facial similarity, in which the similarity of two faces is defined as the Euclidean distance between them in a principal component space, or representational basis. (The space is determined by treating a set of digitized faces as numerical vectors, and by submitting these to principal component analysis). A similar definition is provided for 'facial distinctiveness', namely as the distance of a face from the origin or centroid of the space. The validity of the proposed similarity measure is investigated in several ways, in a total of seven studies, involving approximately 700 subjects. 350 frontal face images and 280 profile face images were collected for use as experimental materials, and as the source for the component space underlying the similarity measure. The weight of the evidence, particularly from a set of similarity rating tasks, suggests that the measure corresponds reasonably well to perceptions of facial similarity. Results from a mock witness experiment showed that it is also strongly, and monotonically related to standard measures of lineup fairness. Evidence from several investigations of the distinctiveness measure, on the other hand, showed that it does not appear to be related to perceptions of facial distinctiveness. An additional empirical investigation examined the relation between target-foil similarity and identification performance. Performance was greater for lineups of low similarity, both when the perpetrator was present, and when the perpetrator was absent. The consequences of this for the understanding of lineup construction and evaluation are discussed.

Machine Learning (ML) is a transformative technology at the forefront of many modern research endeavours. The technology is generating a tremendous amount of attention from researchers and practitioners, providing new approaches to solving complex classification and regression tasks. While concepts such as Deep Learning have existed for many years, the computational power for realising the utility of these algorithms in real-world applications has only recently become available. This dissertation investigated the efficacy of a novel, general method for deploying ML in a variety of complex tasks, where best feature selection, data-set labelling, model definition and training processes were determined automatically. Models were developed in an iterative fashion, evaluated using both training and validation data sets. The proposed method was evaluated using three distinct case studies, describing complex classification tasks often requiring significant input from human experts. The results achieved demonstrate that the proposed method compares with, and often outperforms, less general, comparable methods designed specifically for each task. Feature selection, data-set annotation, model design and training processes were optimised by the method, where less complex, comparatively accurate classifiers with lower dependency on computational power and human expert intervention were produced. In chapter 4, the proposed method demonstrated improved efficacy over comparable systems, automatically identifying and classifying complex application protocols traversing IP networks. In chapter 5, the proposed method was able to discriminate between normal and anomalous traffic, maintaining accuracy in excess of 99%, while reducing false alarms to a mere 0.08%. Finally, in chapter 6, the proposed method discovered more optimal classifiers than those implemented by comparable methods, with classification scores rivalling those achieved by state-of-the-art systems. The findings of this research concluded that developing a fully automated, general method, exhibiting efficacy in a wide variety of complex classification tasks with minimal expert intervention, was possible. The method and various artefacts produced in each case study of this dissertation are thus significant contributions to the field of ML.

No description available.

In the past decades, statistics-based hidden Markov models (HMMs) have become the predominant approach to speech recognition. Under this framework, the speech signal is modeled as a piecewise stationary signal (typically over an interval of 10 milliseconds). Speech features are assumed to be temporally uncorrelated. While these simplifications have enabled tremendous advances in speech processing systems, for the past several years progress on the core statistical models has stagnated. Since machine performance still significantly lags human performance, especially in noisy environments, researchers have been looking beyond the traditional HMM approach. Recent theoretical and experimental studies suggest that exploiting frame-torame correlations in a speech signal further improves the performance of ASR systems. This is typically accomplished by developing an acoustic model which includes higher order statistics or trajectories. Linear Dynamic Models (LDMs) have generated significant interest in recent years due to their ability to model higher order statistics. LDMs use a state space-like formulation that explicitly models the evolution of hidden states using an autoregressive process. This smoothed trajectory model allows the system to better track the speech dynamics in noisy environments. In this dissertation, we develop a hybrid HMM/LDM speech recognizer that effectively integrates these two powerful technologies. This hybrid system is capable of handling large recognition tasks, is robust to noise-corrupted speech data and mitigates the ill-effects of mismatched training and evaluation conditions. This two-pass system leverages the temporal modeling and N-best list generation capabilities of the traditional HMM architecture in a first pass analysis. In the second pass, candidate sentence hypotheses are re-ranked using a phone-based LDM model. The Wall Street Journal (WSJ0) derived Aurora-4 large vocabulary corpus was chosen as the training and evaluation dataset. This corpus is a well-established LVCSR benchmark with six different noisy conditions. The implementation and evaluation of the proposed hybrid HMM/LDM speech recognizer is the major contribution of this dissertation.

No description available.

No description available.