Despite increasing demands for skilled workers within the technological domain, there is still a deficit in the number of graduates in computing fields (computer science, information technology, and computer engineering). Understanding the factors that contribute to students’ motivation and persistence is critical to helping educators, administrators, and industry professionals better focus efforts to improve academic outcomes and job placement. This article examines how experiences contribute to a student’s computing identity, which we define by their interest, recognition, sense of belonging, and competence/performance beliefs. In particular, we consider groups underrepresented in these disciplines, women and minoritized racial/ethnic groups (Black/African American and Hispanic/Latinx). To delve into these relationships, a survey of more than 1,600 students in computing fields was conducted at three metropolitan public universities in Florida. Regression was used to elucidate which experiences predict computing identity and how social identification (i.e., as female, Black/African American, and/or Hispanic/Latinx) may interact with these experiences. Our results suggest that several types of experiences positively predict a student’s computing identity, such as mentoring others, having a job, or having friends in computing. Moreover, certain experiences have a different effect on computing identity for female and Hispanic/Latinx students. More specifically, receiving academic advice from teaching assistants was more positive for female students, receiving advice from industry professionals was more negative for Hispanic/Latinx students, and receiving help on classwork from students in their class was more positive for Hispanic/Latinx students. Other experiences, while having the same effect on computing identity across students, were experienced at significantly different rates by females, Black/African American students, and Hispanic/Latinx students. The findings highlight experiential ways in which computing programs can foster computing identity development, particularly for underrepresented and marginalized groups in computing.

The problem of representative selection amounts to sampling few informative exemplars from large datasets. Existing approaches to data selection often fall short of simultaneously handling non-linear data structures, sampling concise and non-redundant subsets, rejecting outliers, and yielding interpretable outcomes. This paper presents a novel representative selection approach, dubbed MOSAIC, for drawing descriptive sketches of arbitrary manifold structures. Resting upon a novel quadratic formulation, MOSAIC advances a multi-criteria selection approach that maximizes the global representation power of the sampled subset, ensures novelty of the samples by minimizing redundancy, and rejects disruptive information by effectively detecting outliers. Theoretical analyses shed light on geometrical characterization of the obtained sketch and reveal that the sampled representatives maximize a well-defined notion of data coverage in a transformed space. In addition, we present a highly scalable randomized implementation of the proposed algorithm shown to bring about substantial speedups. MOSAIC’s superiority in achieving the desired characteristics of a representative subset all at once while exhibiting remarkable robustness to various outlier types is demonstrated via extensive experiments conducted on both real and synthetic data with comparisons to state-of-the-art algorithms.

Increasingly companies assess a computing candidate's capabilities using technical interviews (TIs). Yet students struggle to code on demand, and there is already an insufficient amount of computing graduates to meet industry needs. Therefore, it is important to understand students' perceptions of TIs, and other professional experiences (e.g., computing jobs). We surveyed 740 undergraduate computing students at three universities to examine their experiences with the hiring process, as well as the impact of professional and cultural experiences (e.g., familial support) on computing identity. We considered the interactions between these experiences and social identity for groups underrepresented in computing - women, Black/African American, and Hispanic/Latinx students. Among other findings, we observed that students that did not have positive experiences with TIs had a reduced computing identity, but that facing discrimination during technical interviews had the opposite effect. Social support may play a role. Having friends in computing bolsters computing identity for Hispanic/Latinx students, as does a supportive home environment for women. Also, freelance computing jobs increase computing identity for Black/African American students. Our findings are intended to raise awareness of the best way for educators to help diverse groups of students to succeed, and to inform them of the experiences that may influence students' engagement, resilience, and computing identity development.

The lack of diversity in computing fields in the United States is a known issue. Students enter the computing fields with the intention of graduating; however, a large number leave and do not persist after enrolling, due to discrimination and biases. This particularly concerns groups already underrepresented in computing fields, such as women, Black/African American students, and Hispanic/Latinx students. However, there are various experiences that can make students feel more included or excluded in the field. Some of these experiences include internships, undergraduate research, capstone courses, and projects, etc. Drawing on Astin's I-E-O model and applying a random forest algorithm, we measure the feature importance of 14 distinct experiences on 1650 students' feelings of inclusivity in the computing field. We observe that there are gender and racial differences in terms of the opinions of computing fields' inclusivity. For example, tutoring experience, job offers, and job experience are considered some of the most important factors for female's perceived inclusiveness of women. However, men perceived women's inclusivity differently, based on the experiences they engaged in. We also looked at the perceived inclusiveness of computing fields for ethnically and racially underrepresented groups, such as Hispanic/Latinx students. Understanding the effect of different experiences on students of both genders with different races and ethnicities on the perceived inclusion could assist the computing community to provide more cohesive experiences that benefits all students and helps them to feel more welcome.

This letter presents an effective and simple method, termed Global Conformity Pursuit (GCP), for robust manifold learning. Data points lying on a union of low-dimensional nonlinear manifolds are expected to be highly conforming. On the other hand, outliers do not typically adhere to low-dimensional structures or otherwise do not exist in large numbers. Hence, they can be identified by their low overall resemblance to the rest of the data. Our kernel-based setting allows us to capture the underlying nonlinear structure of the manifold data, while avoiding the construction of local neighborhood regions which typically causes extreme sensitivity to noise and outliers. Aside from its significantly simple structure-involving only a matrix evaluation and multiplication-GCP is the first manifold learning approach that is simultaneously noniterative and capable of tolerating a large number of outliers, dependent outliers, and noise components. Theoretical analysis guarantees a large gap between conformity values of inliers and outliers. Experimental results showcase potential uses of the proposed framework on benchmark datasets.

This Research Full Paper presents the effects of computing identity sub-constructs on the persistence of computer science students. Computer science (CS) is one of the fastest growing disciplines in the world and an emerging critical field for all students to obtain vital skills to be successful in the 21st century. Despite the growing importance of computer science, many university and college programs suffer from low student persistence rates. Disciplinary identity is a theoretical framework that refers to how students see themselves with respect to a discipline and is related to long-term membership in a disciplinary community. The theory has been effectively applied in Science, Technology, Engineering, and Mathematics (STEM) to understand students' success and persistence. This study examines the effects of performance/competence, recognition, interest and sense of belonging on the academic persistence of computer science students. A survey of approximately 1,640 computing students as part of a National Science Foundation (NSF) funded project was developed and administered at three metropolitan public institutions. Confirmatory Factor Analysis (CFA) was performed to validate the sub-constructs of identity for use in a computing identity model. Then, a structural equation model (SEM) was constructed as a snapshot of the structural relationships for describing and quantifying the impact of the identity sub-constructs on persistence. The results indicated that our model for CS aligns with prior research on disciplinary identity but also adds the importance of sense of belonging. In addition, the findings indicate that students' academic persistence is directly influenced by their interest. A better understanding of these factors may leverage insight into students' academic persistence in computer science/engineering programs as well as a meaningful lens of analysis for further curriculum and extracurricular activities.

In this paper, we introduce a novel hash learning framework for multi-label learning which employs structured prediction. A hash function is learned to embed samples in Hamming spaces, and for each label, a pair of codewords are simultaneously inferred from the available data. These codewords are then used to determine label predictions based on Hamming proximity. The key advantage of this framework is it's computational efficiency in tackling multi-label problems without making restrictive, simplifying assumptions about the structure of the output space, or developing problem-dependent heuristics. Our method not only enjoys considerably better scalability while capturing label inter-dependence, but also yields an exact training algorithm. Experimental results on a collection of benchmark multi-label datasets demonstrate that our model attains higher performance over alternative state-of-the-art multi-label approaches. It is also worth noting that our method can be extended to semi-supervised and missing labels scenarios.

In machine learning, the notion of multi-objective model selection (MOMS) refers to the problem of identifying the set of Pareto-optimal models that optimize by compromising more than one predefined objectives simultaneously. This paper introduces SPRINT-Race, the first multi-objective racing algorithm in a fixed-confidence setting, which is based on the sequential probability ratio with indifference zone test. SPRINT-Race addresses the problem of MOMS with multiple stochastic optimization objectives in the proper Pareto-optimality sense. In SPRINT-Race, a pairwise dominance or non-dominance relationship is statistically inferred via a non-parametric, ternary-decision, dual-sequential probability ratio test. The overall probability of falsely eliminating any Pareto-optimal models or mistakenly returning any clearly dominated models is strictly controlled by a sequential Holm's step-down family-wise error rate control method. As a fixed-confidence model selection algorithm, the objective of SPRINT-Race is to minimize the computational effort required to achieve a prescribed confidence level about the quality of the returned models. The performance of SPRINT-Race is first examined via an artificially constructed MOMS problem with known ground truth. Subsequently, SPRINT-Race is applied on two real-world applications: 1) hybrid recommender system design and 2) multi-criteria stock selection. The experimental results verify that SPRINT-Race is an effective and efficient tool for such MOMS problems. code of SPRINT-Race is available at https://github.com/watera427/SPRINT-Race.

Metric Multi-Dimensional Scaling is one of the most popular visualization techniques aiming to determine a point configuration of the data in a low-dimensional space, while preserving the original pairwise dissimilarities amongst the data points. In this paper we propose Sparse Inductive Multi- Dimensional Scaling approach, an exploratory data analysis and visualization tool. Its kernel-based nature provides an inductive method, and hence allows for embedding of previously unseen data even with non-numerical features. Furthermore, via the use of a sparsity-promoting regularizer, the technique is capable of identifying key training samples, whose contribution in defining the embedded manifold's specific structure is most influential. As we will show, the influence of these samples leads us to interesting interpretations in the context of the data at hand. We adequately train the consequent non-convex framework via an Iterative Majorization approach, for which a stable closed-form solution is obtained. Experimental results showcase the potential benefits of the proposed embedding method on both artificial and real-world data.

Model selection is a core aspect in machine learning and is, occasionally, multi-objective in nature. For instance, hyper-parameter selection in a multi-task learning context is of multi-objective nature, since all the tasks' objectives must be optimized simultaneously. In this paper, a novel multi-objective racing algorithm (RA), namely S-Race, is put forward. Given an ensemble of models, our task is to reliably identify Pareto optimal models evaluated against multiple objectives, while minimizing the total computational cost. As a RA, S-Race attempts to eliminate non-promising models with confidence as early as possible, so as to concentrate computational resources on promising ones. Simultaneously, it addresses the problem of multi-objective model selection (MOMS) in the sense of Pareto optimality. In S-Race, the nonparametric sign test is utilized for pair-wise dominance relationship identification. Moreover, a discrete Holm's step-down procedure is adopted to control the family-wise error rate of the set of hypotheses made simultaneously. The significance level assigned to each family is adjusted adaptively during the race. In order to illustrate its merits, S-Race is applied on three MOMS problems: 1) selecting support vector machines for classification; 2) tuning the parameters of artificial bee colony algorithms for numerical optimization; and 3) constructing optimal hybrid recommendation systems for movie recommendation. The experimental results confirm that S-Race is an efficient and effective MOMS algorithm compared to a brute-force approach.

The support vector machine (SVM) remains a popular classifier for its excellent generalization performance and applicability of kernel methods; however, it still requires tuning of a regularization parameter, C , to achieve optimal performance. Regularization path-following algorithms efficiently solve the solution at all possible values of the regularization parameter relying on the fact that the SVM solution is piece-wise linear in C . The SVMPath originally introduced by Hastie et al., while representing a significant theoretical contribution, does not work with semidefinite kernels. Ong et al. introduce a method improved SVMPath (ISVMP) algorithm, which addresses the semidefinite kernel; however, Singular Value Decomposition or QR factorizations are required, and a linear programming solver is required to find the next C value at each iteration. We introduce a simple implementation of the path-following algorithm that automatically handles semidefinite kernels without requiring a method to detect singular matrices nor requiring specialized factorizations or an external solver. We provide theoretical results showing how this method resolves issues associated with the semidefinite kernel as well as discuss, in detail, the potential sources of degeneracy and cycling and how cycling is resolved. Moreover, we introduce an initialization method for unequal class sizes based upon artificial variables that work within the context of the existing path-following algorithm and do not require an external solver. Experiments compare performance with the ISVMP algorithm introduced by Ong et al. and show that the proposed method is competitive in terms of training time while also maintaining high accuracy.

Multi-objective model selection, which is an important aspect of Machine Learning, refers to the problem of identifying a set of Pareto optimal models from a given ensemble of models. This paper proposes SPRINT-Race, a multi-objective racing algorithm based on the Sequential Probability Ratio Test with an Indifference Zone. In SPRINT-Race, a non-parametric ternary-decision sequential analogue of the sign test is adopted to identify pair-wise dominance and non-dominance relationship. In addition, a Bonferroni approach is employed to control the overall probability of any erroneous decisions. In the fixed confidence setting, SPRINT-Race tries to minimize the computational effort needed to achieve a predefined confidence about the quality of the returned models. The efficiency of SPRINT-Race is analyzed on artificially-constructed multi-objective model selection problems with known ground-truth. Moreover, SPRINT-Race is applied to identifying the Pareto optimal parameter settings of Ant Colony Optimization algorithms in the context of solving Traveling Salesman Problems. The experimental results confirm the advantages of SPRINT-Race for multi-objective model selection.

A traditional and intuitively appealing Multitask Multiple Kernel Learning (MT-MKL) method is to optimize the sum (thus, the average) of objective functions with (partially) shared kernel function, which allows information sharing among the tasks. We point out that the obtained solution corresponds to a single point on the Pareto Front (PF) of a multiobjective optimization problem, which considers the concurrent optimization of all task objectives involved in the Multitask Learning (MTL) problem. Motivated by this last observation and arguing that the former approach is heuristic, we propose a novel support vector machine MT-MKL framework that considers an implicitly defined set of conic combinations of task objectives. We show that solving our framework produces solutions along a path on the aforementioned PF and that it subsumes the optimization of the average of objective functions as a special case. Using the algorithms we derived, we demonstrate through a series of experimental results that the framework is capable of achieving a better classification performance, when compared with other similar MTL approaches.

This paper presents a pair of hypothesis spaces (HSs) of vector-valued functions intended to be used in the context of multitask classification. While both are parameterized on the elements of reproducing kernel Hilbert spaces and impose a feature mapping that is common to all tasks, one of them assumes this mapping as fixed, while the more general one learns the mapping via multiple kernel learning. For these new HSs, empirical Rademacher complexity-based generalization bounds are derived, and are shown to be tighter than the bound of a particular HS, which has appeared recently in the literature, leading to improved performance. As a matter of fact, the latter HS is shown to be a special case of ours. Based on an equivalence to Group-Lasso type HSs, the proposed HSs are utilized toward corresponding support vector machine-based formulations. Finally, experimental results on multitask learning problems underline the quality of the derived bounds and validate this paper's analysis.

Racing algorithms are often used for offline model selection, where models are compared in terms of their average performance over a collection of problems. In this paper, we present a new racing algorithm variant, Max-Race, which makes decisions based on the maximum performance of models. It is an online algorithm, whose goal is to optimally allocate computational resources in a portfolio of evolutionary algorithms, while solving a particular problem instance. It employs a hypothesis test based on extreme value theory in order to decide, which component algorithms to retire, while avoiding unnecessary computations. Experimental results confirm that Max-Race is able to identify the best individual with high precision and low computational overhead. When used as a scheme to select the best from a portfolio of algorithms, the results compare favorably to the ones of other popular algorithm portfolio approaches.

Multi-task learning (MTL) is an established method of inducing bias in neural network learning. It has been used as a method of knowledge transfer in sequential learning tasks, and as a method of overcoming the problem of impoverished training data. A standard problem in neural networks is the establishment of the proper network structure for a given problem. This problem is compounded in the MTL case since the structure of the network can significantly impact the effectiveness of the MTL method. This problem is exacerbated further in the sequential or life-long learning case, as choosing structure a priori for an unknown number of tasks is not possible. In this paper, we explore the use of cascade correlation (CC) as a generative neural network algorithm to overcome these limitations in MTL. We show that generative neural networks using the CC algorithm can display the same MTL advantages while alleviating the need to choose the network structure a priori. We then incorporate the functional transfer from the MTL techniques with the generative structure of CC networks to build a dual network life-long learning system that does not require the network topology or the number of tasks to be predefined.

Over the past few years, multiple kernel learning (MKL) has received significant attention among data-driven feature selection techniques in the context of kernel-based learning. MKL formulations have been devised and solved for a broad spectrum of machine learning problems, including multitask learning (MTL). Solving different MKL formulations usually involves designing algorithms that are tailored to the problem at hand, which is, typically, a nontrivial accomplishment. In this paper we present a general multitask multiple kernel learning (MT-MKL) framework that subsumes well-known MT-MKL formulations, as well as several important MKL approaches on single-task problems. We then derive a simple algorithm that can solve the unifying framework. To demonstrate the flexibility of the proposed framework, we formulate a new learning problem, namely partially-shared common space MT-MKL, and demonstrate its merits through experimentation.

This paper presents a multi-objective racing algorithm, S-Race, which efficiently addresses multi-objective model selection problems in the sense of Pareto optimality. As a racing algorithm, S-Race attempts to eliminate candidate models as soon as there is sufficient statistical evidence of their inferiority relative to other models with respect to all objectives. This approach is followed in the interest of controlling the computational effort. S-Race adopts a non-parametric sign test to identify pair-wise domination relationship between models. Meanwhile, Holm's Step-Down method is employed to control the overall family-wise error rate of simultaneous hypotheses testing during the race. Experimental results involving the selection of superior Support Vector Machine classifiers according to 2 and 3 performance criteria indicate that S-Race is an efficient and effective algorithm for automatic model selection, when compared to a brute-force, multi-objective selection approach.

In this paper we present two related, kernel-based Distance Metric Learning (DML) methods. Their respective models non-linearly map data from their original space to an output space, and subsequent distance measurements are performed in the output space via a Mahalanobis metric. The dimensionality of the output space can be directly controlled to facilitate the learning of a low-rank metric. Both methods allow for simultaneous inference of the associated metric and the mapping to the output space, which can be used to visualize the data, when the output space is 2-or 3-dimensional. Experimental results for a collection of classification tasks illustrate the advantages of the proposed methods over other traditional and kernel-based DML approaches.

Research studies have shown the importance of the relationship between “a sense of belonging” and successfully adjusting to college life for undergraduate students. Few studies, however, have compared if native freshmen and transfer students connect to the campus in similar ways. The purpose of this study was to examine the relationships among participation in social and academic engagement activities, social integration, and academic success for students graduating with engineering or computer science Bachelor's degrees at a large, public university. An additional area of interest was the role of undergraduate research experiences (UREs) in making friends on campus. Using analysis of variance, 1,042 responses were analyzed from senior exit survey questions and cumulative GPA from student records. Students who reported higher levels of social integration on average had higher GPAs. Seven activities were found to be strongly associated with social integration for native freshmen, and three activities for transfer students. Three activities - organizations in their major, experiential learning, and intramural sports - were common to both groups. Surprisingly, given the cohort experience structure of UREs, no significant relationship was found for social integration in this study. This finding may be limited by the sample size or other factors such as self-selection bias.

The Sammon Mapping (SM) has established itself as a valuable tool in dimensionality reduction, manifold learning, exploratory data analysis and, particularly, in data visualization. The SM is capable of projecting high-dimensional data into a low-dimensional space, so that they can be visualized and interpreted. This is accomplished by representing inter-sample dissimilarities in the original space by Euclidean inter-sample distances in the projection space. Recently, Kernel Sammon Mapping (KSM) has been shown to subsume the SM and a few other related extensions to SM. Both of the aforementioned models feature a set of linear weights that are estimated via Iterative Majorization (IM). While IM is significantly faster than other standard gradient-based methods, tackling data sets of larger than moderate sizes becomes a challenging learning task, as IM's convergence significantly slows down with increasing data set cardinality. In this paper we derive two improved training algorithms based on Successive Over-Relaxation (SOR) and Parallel Tangents (PARTAN) acceleration, that, while still being first-order methods, exhibit faster convergence than IM. Both algorithms are relatively easy to understand, straightforward to implement and, performance-wise, are as robust as IM. We also present comparative results that illustrate their computational advantages on a set of benchmark problems.

Over the last few years, Kernel Principal Component Analysis (KPCA) has found several applications in outlier detection. A relatively recent method uses KPCA to compute the reconstruction error (RE) of previously unseen samples and, via thresholding, to identify atypical samples. In this paper we propose an alternative method, which performs the same task, but considers Mahalanobis distances in the orthogonal complement of the subspace that is utilized to compute the reconstruction error. In order to illustrate its merits, we provide qualitative and quantitative results on both artificial and real datasets and we show that it is competitive, if not superior, for several outlier detection tasks, when compared to the original RE-based variant and the One-Class SVM detection approach.

In this paper the major principles to effectively design a parameter-less, multi-objective evolutionary algorithm that optimizes a population of probabilistic neural network (PNN) classifier models are articulated; PNN is an example of an exemplar-based classifier. These design principles are extracted from experiences, discussed in this paper, which guided the creation of the parameter-less multi-objective evolutionary algorithm, named MO-EPNN (multi-objective evolutionary probabilistic neural network). Furthermore, these design principles are also corroborated by similar principles used for an earlier design of a parameter-less, multi-objective genetic algorithm used to optimize a population of ART (adaptive resonance theory) models, named MO-GART (multi-objective genetically optimized ART); the ART classifier model is another example of an exemplar-based classifier model. MO-EPNN's performance is compared to other popular classifier models, such as SVM (Support Vector Machines) and CART (Classification and Regression Trees), as well as to an alternate competitive method to genetically optimize the PNN. These comparisons indicate that MO-EPNN's performance (generalization on unseen data and size) compares favorably to the aforementioned classifier models and to the alternate genetically optimized PNN approach. MO-EPPN's good performance, and MO-GART's earlier reported good performance, both of whose design relies on the same principles, gives credence to these design principles, delineated in this paper.

Existing active set methods reported in the literature for support vector machine (SVM) training must contend with singularities when solving for the search direction. When a singularity is encountered, an infinite descent direction can be carefully chosen that avoids cycling and allows the algorithm to converge. However, the algorithm implementation is likely to be more complex and less computationally efficient than would otherwise be required for an algorithm that does not have to contend with the singularities. We show that the revised simplex method introduced by Rusin provides a guarantee of nonsingularity when solving for the search direction. This method provides for a simpler and more computationally efficient implementation, as it avoids the need to test for rank degeneracies and also the need to modify factorizations or solution methods based upon those rank degeneracies. In our approach, we take advantage of the guarantee of nonsingularity by implementing an efficient method for solving the search direction and show that our algorithm is competitive with SVM-QP and also that it is a particularly effective when the fraction of nonbound support vectors is large. In addition, we show competitive performance of the proposed algorithm against two popular SVM training algorithms, SVMLight and LIBSVM.

In this paper, we present the evolution of adaptive resonance theory (ART) neural network architectures (classifiers) using a multiobjective optimization approach. In particular, we propose the use of a multiobjective evolutionary approach to simultaneously evolve the weights and the topology of three well-known ART architectures; fuzzy ARTMAP (FAM), ellipsoidal ARTMAP (EAM), and Gaussian ARTMAP (GAM). We refer to the resulting architectures as MO-GFAM, MO-GEAM, and MO-GGAM, and collectively as MO-GART. The major advantage of MO-GART is that it produces a number of solutions for the classification problem at hand that have different levels of merit [accuracy on unseen data (generalization) and size (number of categories created)]. MO-GART is shown to be more elegant (does not require user intervention to define the network parameters), more effective (of better accuracy and smaller size), and more efficient (faster to produce the solution networks) than other ART neural network architectures that have appeared in the literature. Furthermore, MO-GART is shown to be competitive with other popular classifiers, such as classification and regression tree (CART) and support vector machines (SVMs).

Efficiently implemented active set methods have been successfully applied to support vector machine (SVM) training. These active set methods offer higher precision and incremental training at the cost of additional memory requirements when compared to decomposition methods such as sequential minimal optimization (SMO). However, all existing active set methods must deal with singularities occurring within the inner problem solved at each iteration, a problem that leads to more complex implementation and potential inefficiencies. Here, we introduce a revised simplex method, originally introduced by Rusin, adapted for SVM training and show this is an active set method similar to most existing methods with the advantage of maintaining nonsingularity of the inner problem. We compare performance to an existing active set method introduced by Scheinberg and demonstrate an improvement in training times, in some cases. We show our method maintains a slightly simpler implementation and offers advantages in terms of applying iterative methods to alleviate memory concerns. We also show performance of the active set methods when compared to state-of-the-art decomposition implementations such as SVMLight and SMO.

In this work we present, for the first time, the evolution of ART Neural Network architectures (classifiers) using a multiobjective optimization approach. In particular, we propose the use of a multiobjective evolutionary approach to evolve simultaneously the weights, as well as the topology of three well-known ART architectures; Fuzzy ARTMAP (FAM), Ellipsoidal ARTMAP (EAM) and Gaussian ARTMAP (GAM). We refer to the resulting architectures as MO-GFAM, MOGEAM, or MO-GGAM, and collectively as MO-GART. The major advantage of MO-GART is that it produces a number of solutions for the classification problem at hand that have different levels of merit (accuracy on unseen data (generalization) and size (number of categories created)). MO-GART is shown to be more elegant (does not require user intervention to define the network parameters), more effective (of better accuracy and smaller size), and more efficient (faster to produce the solution networks) than other ART neural network architectures that have appeared in the literature.

Genetic algorithms have been used to evolve several neural network architectures. In a previous effort, we introduced the evolution of three well known ART architects; Fuzzy ARTMAP (FAM), Ellipsoidal ARTMAP (EAM) and Gaussian ARTMAP (GAM). The resulting architectures were shown to achieve competitive generalization and exceptionally small size. A major concern regarding these architectures, and any evolved neural network architecture in general, is the added overhead in terms of computational time needed to produce the finally evolved network. In this paper we investigate ways of reducing this computational overhead by reducing the computations needed for the calculation of the fitness value of the evolved ART architectures. The results obtained in this paper can be directly extended to many other evolutionary neural network architectures, beyond the studied evolution of ART neural network architectures.

Active set methods for training the support vector machines (SVM) are advantageous since they enable incremental training and, as we show in this research, do not exhibit exponentially increasing training times commonly associated with the decomposition methods as the SVM training parameter, C, is increased or the classification difficulty increases. Previous implementations of the active set method must contend with singularities, especially associated with the linear kernel, and must compute infinite descent directions, which may be inefficient, especially as C is increased. In this research, we propose a revised simplex method for quadratic programming, which has a guarantee of non-singularity for the sub-problem, and show how this can be adapted to SVM training.

Pruning is one of the key procedures in training decision tree classifiers. It removes trivial rules from the raw knowledge base built from training examples, in order to avoid over-using noisy, conflicting, or fuzzy inputs, so that the refined model can generalize better with unseen cases. In this paper, we present a number of properties of k-norm pruning, a recently proposed pruning algorithm, which has clear theoretical interpretation. In an earlier paper it was shown that k-norm pruning compares very favorably in terms of accuracy and size with minimal cost-complexity pruning and error based pruning, two of the most cited decision tree pruning methods; it was also shown that k-norm pruning is more efficient, at times orders of magnitude more efficient than minimal cost-complexity pruning and error based pruning. In this paper, we demonstrate the validity of the k-norm properties through a series of theorems, and explain their practical significance.

Outlier detection has received considerable attention in many applications, such as detecting network attacks or credit card fraud The massive datasets currently available for mining in some of these outlier detection applications require large parallel systems, and consequently parallelizable outlier detection methods. Most existing outlier detection methods assume that all of the attributes of a dataset are numerical, usually have a quadratic time complexity with respect to the number of points in the dataset, and quite often they require multiple dataset scans. In this paper, we propose a fast parallel outlier detection strategy based on the Attribute Value Frequency (AVF) approach, a high-speed, scalable outlier detection method for categorical data that is inherently easy to parallelize. Our proposed solution, MR-AVF, is based on the MapReduce paradigm for parallel programming, which offers load balancing and fault tolerance. MR-AVF is particularly simple to develop and it is shown to be highly scalable with respect to the number of cluster nodes.

Outlier detection has received significant attention in many applications, such as detecting credit card fraud or network intrusions. Most existing research focuses on numerical datasets, and cannot directly apply to categorical sets where there is little sense in calculating distances among data points. Furthermore, a number of outlier detection methods require quadratic time with respect to the dataset size and usually multiple dataset scans. These characteristics are undesirable for large datasets, potentially scattered over multiple distributed sites. In this paper, we introduce Attribute Value Frequency (A VF), a fast and scalable outlier detection strategy for categorical data. A VF scales linearly with the number of data points and attributes, and relies on a single data scan. AVF is compared with a list of representative outlier detection approaches that have not been contrasted against each other. Our proposed solution is experimentally shown to be significantly faster, and as effective in discovering outliers.

In the process of learning a pattern I, the fuzzy ARTMAP algorithm templates (i.e., the weight vectors corresponding to nodes of its category representation layer) compete for the representation of the given pattern. This competition can inducematchtracking: a process that iterates a number of times over the template set searching for a templatew*of the correct class that best represents the pattern I. In this paper, we analyze the search for a winning template from the perspective of bi-criterion optimization and prove that it is actually a walk along the Pareto front of an appropriately defined co-objective optimization problem. This observation allows us to propose the basis for an implementation variant of Fuzzy ARTMAP that (a) produces exactly the same network as Fuzzy ARTMAP, (b) avoids matchtracking by explicitly keeping track of a subset of the Pareto front, (c) finds the correct template to represent an input pattern through a single pass over the template set and (d) eliminates the need for the Fuzzy ARTMAP parameter ε.

This paper focuses on the evolution of ARTMAP architectures, using genetic algorithms, with the objective of improving generalization performance and alleviating the ART category proliferation problem. We refer to the resulting architectures as GFAM, GEAM, and GGAM. We demonstrate through extensive experimentation that evolved ARTMAP architectures exhibit good generalization and are of small size, while consuming reasonable computational effort to produce an optimal or a sub-optimal network. Furthermore, we compare the performance of GFAM, GEAM and GGAM with other competitive ARTMAP architectures that have appeared in the literature and addressed the category proliferation problem in ART. This comparison indicates that GFAM, GEAM and GGAM have superior performance (generalize better, are of smaller size, and require less computations) compared with other competitive ARTMAP architectures.

In this paper, we present a version of the probabilistic neural network (PNN) that is capable of operating on a distributed database that is horizontally partitioned. It does so in a way that is privacy-preserving: that is, a test point can be evaluated by the algorithm without any party knowing the data owned by the other parties. We present an analysis of this algorithm from the standpoints of security and computational performance. Finally, we provide performance results of an implementation of this privacy preserving, distributed PNN algorithm.

The support vector machine is a widely employed machine learning model due to its repeatedly demonstrated superior generalization performance. The sequential minimal optimization (SMO) algorithm is one of the most popular SVM training approaches. SMO is fast, as well as easy to implement; however, it has a limited working set size (2 points only). Faster training times can result if the working set size can be increased without significantly increasing the computational complexity. In this paper, we extend the 2-point SMO formulation to a 4-point formulation and address the theoretical issues associated with such an extension. We show that modifying the SMO algorithm to increase the working set size is beneficial in terms of the number of iterations required for convergence, and shows promise for reducing the overall training time.

Probabilistic neural networks (PNN) and general regression neural networks (GRNN) represent knowledge by simple but interpretable models that approximate the optimal classifier or predictor in the sense of expected value of the accuracy. These models require the specification of an important smoothing parameter, which is usually chosen by cross-validation or clustering. In this article, we demonstrate the problems with the cross-validation and clustering approaches to specify the smoothing parameter, discuss the relationship between this parameter and some of the data statistics, and attempt to develop a fast approach to determine the optimal value of this parameter. Finally, through experimentation, we show that our approach, referred to as a gap-based estimation approach, is superior in speed to the compared approaches, including support vector machine, and yields good and stable accuracy.

Fuzzy ARTMAP (FAM) is currently considered to be one of the premier neural network architectures in solving classification problems. One of the limitations of fuzzy ARTMAP that has been extensively reported in the literature is the category proliferation problem. That is fuzzy ARTMAP has the tendency of increasing its network size, as it is confronted with more and more data, especially if the data is of noisy and/or overlapping nature. To remedy this problem a number of researchers have designed modifications to the training phase of fuzzy ARTMAP that had the beneficial effect of reducing this phenomenon. In this paper we propose a new approach to handle the category proliferation problem in fuzzy ARTMAP by evolving trained FAM architectures. We refer to the resulting FAM architectures as GFAM We demonstrate through extensive experimentation that an evolved FAM (GFAM) exhibits good generalization, small size, and produces an optimal or a good sub-optimal network with a reasonable computational effort. Furthermore, comparisons of the GFAM with other approaches, proposed in the literature, that address the FAM category proliferation problem, illustrate that the GFAM has a number of advantages (i.e. produces smaller or equal size architectures, of better or as good generalization, with reduced computational complexity).

In this paper, we present three different methods for implementing the Probabilistic Neural Network on a Beowulf cluster computer. The three methods, Parallel Full Training Set (PFT-PNN), Parallel Split Training Set (PST-PNN) and the Pipelined PNN (PPNN) all present different performance tradeoffs for different applications. We present implementations for all three architectures that are fully equivalent to the serial version and analyze the tradeoffs governing their potential use in actual engineering applications. Finally we provide performance results for all three methods on a Beowulf cluster.

Probabilistic neural networks (PNN) and general regression neural networks (GRNN) represent the knowledge by a simple but interpretable model that approximates the optimal classifier/predictor in the sense of expected value of accuracy. This model requires an important preset smoothing parameter, which is usually chosen by cross-validation or clustering. In this paper, we demonstrate the difficulties of both these approaches, discuss the relationship between this parameter and some of the data statistics, and attempt to develop a fast approach to determine the optimal value of this parameter. Finally, through experimentation we show that our approach, referred to as a gap-based estimation approach, is superior to the compared approaches.

Fuzzy ARTMAP (FAM) is currently considered as one of the premier neural network architectures in solving classification problems. Safe ARTMAP, a modified version of FAM, was introduced to remedy the category proliferation problem that has been extensively reported in the literature. However, Safe ARTMAP's performance depends on a number of parameters. In this paper, we analyzed each parameter to set up the candidate values for evaluation. We performed an exhaustive experimentation to identify good default values for these parameters for a variety of problems, and compared the best performing Safe ARTMAP network with other best performing ART networks, including those that claim to solve the category proliferation problem.

It is widely accepted that the difficulty and expense involved in acquiring the knowledge behind tactical behaviors has been one limiting factor in the development of simulated agents representing adversaries and teammates in military and game simulations. Several researchers have addressed this problem with varying degrees of success. The problem mostly lies in the fact that tactical knowledge is difficult to elicit and represent through interactive sessions between the model developer and the subject matter expert. This paper describes a novel approach that employs genetic programming in conjunction with context-based reasoning to evolve tactical agents based upon automatic observation of a human performing a mission on a simulator. In this paper, we describe the process used to carry out the learning. A prototype was built to demonstrate feasibility and it is described herein. The prototype was rigorously and extensively tested. The evolved agents exhibited good fidelity to the observed human performance, as well as the capacity to generalize from it.

In this paper we present an experimental comparison of four neural-based classifiers, namely growing cell structures (GCS), growing neural gas (GNG), semi-supervised fuzzy ARTMAP (ssFAM) and semi-supervised ellipsoid ARTMAP (ssEAM). The comparison is performed in terms of classification accuracy and structural complexity of the resulting classifiers. Earlier studies that had appeared in the literature showed that fuzzy ARTMAP, which utilizes fully-supervised learning, may suffer from poor generalization performance, when compared to GCS and GNG classifiers. This phenomenon typically occurs, when class distribution overlap is significant. Here, we present new results indicating that ARTMAP classifiers equipped with semi-supervised learning capabilities can improve their performance with respect to GCS and GNG classifiers, while maintaining lower structural complexity.

Fuzzy ARTMAP neural networks have been proven to be good classifiers on a variety of classification problems. However, the time that it takes fuzzy ARTMAP to converge to a solution increases rapidly as the number of patterns used for training increases. In this paper, we propose a coarse grain parallelization technique, based on a pipeline approach, to speed-up fuzzy ARTMAP's training process. In particular, we first parallelized fuzzy ARTMAP, without the match-tracking mechanism, and then we parallelized fuzzy ARTMAP with the match-tracking mechanism. Results run on a Beowulf cluster with a well known large database (Forrest Covertype database from the UCI repository) show linear speedup with respect to the number of processors used in the pipeline.

One of the properties of FAM, which is a mixed blessing, is its capacity to produce new neurons (templates) on demand to represent classification categories. This property allows FAM to automatically adapt to the database without having to arbitrarily specify network structure, but it also has the undesirable side effect that on large databases it can produce a large network size that can dramatically slow down the algorithms' training time. To address this problem, we propose the use of the Hilbert space-filling curve. Our results indicate that the Hilbert space-filling curve can reduce the training time of FAM by partitioning the learning set without a significant effect on the classification performance or network size. Given that there is full data partitioning with the HSFC, we implement and test a parallel implementation on a Beowulf cluster of workstations that further speeds up the training and classification time on large databases.

The focus of this paper is semi-supervised learning in the context of pattern recognition. Semi-supervised learning (SSL) refers to the semi-supervised construction of clusters during the training phase of exemplar-based classifiers. Using artificially generated data sets we present experimental results of classifiers that follow the SSL paradigm and we show that, especially for difficult pattern recognition problems featuring high class overlap, for exemplar-based classifiers implementing SSL i) the generalization performance improves, while ii) the number of necessary exemplars decreases significantly, when compared to the original versions of the classifiers.

A measure of success for any learning algorithm is how useful it is in a variety of learning situations. Those learning algorithms that support universal function approximation can theoretically be applied to a very large and interesting class of learning problems. Many kinds of neural network architectures have already been shown to support universal approximation. In this paper, we will provide a proof to show that Fuzzy ART augmented with a single layer of perceptrons is a universal approximator. Moreover, the Fuzzy ARTMAP neural network architecture, by itself, will be shown to be a universal approximator.

This paper describes an approach to classification of noisy signals using a technique based on the fuzzy ARTMAP neural network (FAMNN). The proposed method is a modification of the testing phase of the fuzzy ARTMAP that exhibits superior generalization performance compared to the generalization performance of the standard fuzzy ARTMAP in the presence of noise. An application to textured gray-scale image segmentation is presented. The superiority of the proposed modification over the standard fuzzy ARTMAP is established by a number of experiments using various texture sets, feature vectors and noise types. The texture sets include various aerial photos and also samples obtained from the Brodatz album. Furthermore, the classification performance of the standard and the modified fuzzy ARTMAP is compared for different network sizes. Classification results that illustrate the performance of the modified algorithm and the FAMNN are presented.

A novel adaptive resonance theory (ART) neural network architecture is being proposed. The new model, called Hypersphere ART (H-ART) is based on the same principals as Fuzzy-ART and, thus, inherits most of its qualities for unsupervised learning. Among these properties is fast, stable, incremental learning on the training set and good generalization on the testing set. While H-ART is intended for clustering tasks, its extension, H-ARTMAP is playing the role of Fuzzy-ARTMAP's counterpart for the supervised learning of real-valued, multi-dimensional mappings. Also in this paper, some experimental results are presented involving the comparison of H-ARTMAP and Fuzzy-ARTMAP in simple, illustrative classification problems. The results are indicating comparable performances in error rate but also a good potential for substantial superiority of H-ARTMAP in terms of nodes (categories) utilized. The latter effect can be attributed to H-ART's more efficient internal knowledge representation.

This paper describes an approach to classification of noisy signals using a technique based on the Fuzzy ARTMAP neural network (FAM). A variation of the testing phase of Fuzzy ARTMAP is introduced, that exhibited superior generalization performance than the standard Fuzzy ARTMAP in the presence of noise. We present an application of our technique for textured grayscale images. We perform a large number of experiments to verify the superiority of the modified over the standard Fuzzy ARTMAP. More specifically, the modified and the standard FAM were evaluated on two different sets of features (fractal-based and energy-based), for three different types of noise (Gaussian, uniform, exponential) and for two different texture sets (Brodatz, aerial). Furthermore, the classification performance of the standard and modified Fuzzy ARTMAP was compared for different network sizes.

We present a modification to the Fuzzy ARTMAP neural network architecture for conducting classification in a probabilistic setting. We call this new architecture Hierarchical ARTMAP (HARTMAP). Performance comparisons with Fuzzy ARTMAP, Gaussian ARTMAP and Boosted ARTMAP on some simple two-class problems are discussed. Experimental results indicate that HARTMAP yields better generalization results on problems involving overlap of the underlying pattern distributions.

In this paper we introduce a procedure, based on the max-min clustering method, that identifies a fixed order of training pattern presentation for fuzzy adaptive resonance theory mapping (ARTMAP). This procedure is referred to as the ordering algorithm, and the combination of this procedure with fuzzy ARTMAP is referred to as ordered fuzzy ARTMAP. Experimental results demonstrate that ordered fuzzy ARTMAP exhibits a generalization performance that is better than the average generalization performance of fuzzy ARTMAP, and in certain cases as good as, or better than the best fuzzy ARTMAP generalization performance. We also calculate the number of operations required by the ordering algorithm and compare it to the number of operations required by the training phase of fuzzy ARTMAP. We show that, under mild assumptions, the number of operations required by the ordering algorithm is a fraction of the number of operations required by fuzzy ARTMAP.

Current indexing-based approaches build the hash table using either a large number of reference views or 3D models. In this paper, we propose building the hash table using algebraic functions of views. During preprocessing, we consider groups of model points and we represent all the views (i.e., images) that they can produce in a hash table. These views are computed using algebraic functions of a small number of reference views which contain the group. Fundamental to this procedure is a methodology based on singular value decomposition and interval arithmetic for estimating the ranges of values that the parameters of algebraic functions can assume. During recognition, scene groups are used to retrieve from the hash table the model groups that might have produced them. Using algebraic functions of views for indexing-based recognition offers a number of advantages. First of all, the hash table can be built easier without requiring 3D models or a large number of reference views. Second, recognition does not rely on the similarities between new and reference views. Third, verification becomes simpler. Finally, the approach is more general and extendible.

A major problem associated with geometric hashing and methods which have emerged from it is the nonuniform distribution of invariants over the hash space. This has two serious effects on the performance of the method. First, it can result in an inefficient storage of data which can increase recognition time. Second, given that geometric hashing is highly amenable to parallel implementation, a nonuniform distribution of data poses difficulties in tackling the load-balancing problem. Finding a "good" geometric hash function which redistributes the invariants uniformly over the hash space is not easy. Current approaches make assumptions about the statistical characteristics of the data and then use techniques from probability theory to calculate a transformation that maps the nonuniform distribution of invariants to a uniform one. In this paper, a new approach is proposed based on an elastic hash table. In contrast to existing approaches which try to redistribute the invariants over the hash bins, we proceed oppositely by distributing the hash bins over the invariants. The key idea is to associate the hash bins with the output nodes of a self-organizing feature map (SOFM) neural network which is trained using the invariants as training examples. In this way, the location of a hash bin in the space of invariants is determined by the weight vector of the node associated with the hash bin. During training, the SOFM spreads the hash bins proportionally to the distribution of invariants (i.e., more hash bins are assigned to higher density areas while less hash bins are assigned to lower density areas) and adjusts their size so that they eventually hold almost the same number of invariants. The advantage of the proposed approach is that it is a process that adapts to the invariants through learning. Hence, it makes absolutely no assumptions about the statistical characteristics of the invariants and the geometric hash function is actually computed through learning. Furthermore, SOFM's "topology preserving" property ensures that the computed geometric hash function should be well behaved. The proposed approach, was shown to perform well on both artificial and real data.

We consider the problem of learning the mapping between the image coordinates of unknown affine views of an object and the parameters of the affine transformation that can align a known view of the same object with them. A single layer neural network (SL-NN) is used to learn the mapping. Although the proposed approach is conceptually similar to other approaches in the literature, its practical advantages are more profound. The views used to train the SL-NN are not obtained by taking different pictures of the object but by sampling the space of its affine transformed views. This space is constructed by estimating the range of values that the parameters of affine transformation can assume using a single view and a methodology based on singular value decomposition and interval arithmetic. The proposed scheme is as accurate as traditional least-squares approaches but faster. A front-end stage to the SL-NN, based on principal components analysis increases its noise tolerance dramatically and guides us in deciding how many training views are necessary in order for it to learn a good mapping.

A major problem associated with geometric hashing and methods which have emerged from it is the non-uniform distribution of invariants over the hash space. This problem can affect the performance of the method significantly. Finding a "good" geometric hash function which redistributes the invariants uniformly over the hash space is not easy. In this paper, a new approach is proposed for alleviating the above problem. It is based on the use of an "elastic hash table" which is implemented as a self-organizing feature map neural network (SOFM-NN). In contrast to existing approaches which try to redistribute the invariants over the hash bins, we proceed oppositely, spreading the hash bins over the invariants. During training, the SOFM-NN resembles an elastic net which deforms over the hash space. The objective of the deformation process is to spread more hash bins in hash space areas which are heavily occupied and less hash bins in lower density areas. The advantage of the proposed approach is that it is a process that adapts to the invariants through learning. Hence, it makes absolutely no assumptions about the statistical characteristics of the invariants and the geometric hash function is actually computed through learning. Furthermore, the well known "topology preserving" property of the SOFM-NN guarantees that the computed geometric hash function should be well behaved. Finally, the proposed approach is inherently parallelizable.>

A set of nonlinear differential equations that describe the dynamics of the ART1 model are presented, along with the motivation for their use. These equations are extensions of those developed by Carpenter and Grossberg (1987). It is shown how these differential equations allow the ART1 model to be realized as a collective nonlinear dynamical system. Specifically, we present an ART1-based neural network model whose description requires no external control features. That is, the dynamics of the model are completely determined by the set of coupled differential equations that comprise the model. It is shown analytically how the parameters of this model can be selected so as to guarantee a behavior equivalent to that of ART1 in both fast and slow learning scenarios. Simulations are performed in which the trajectories of node and weight activities are determined using numerical approximation techniques.

The analysis of complex neural network models via analytical techniques is often quite difficult due to the large numbers of components involved and the nonlinearities associated with these components. The authors present a framework for simulating neural networks as discrete event nonlinear dynamical systems. This includes neural network models whose components are described by continuous-time differential equations or by discrete-time difference equations. Specifically, the authors consider the design and construction of a concurrent object-oriented discrete event simulation environment for neural networks. The use of an object-oriented language provides the data abstraction facilities necessary to support modification and extension of the simulation system at a high level of abstraction. Furthermore, the ability to specify concurrent processing supports execution on parallel architectures. The use of this system is demonstrated by simulating a specific neural network model on a general-purpose parallel computer.>

The author corrects the approach followed in the above-titled paper (see ibid., vol.36, no.6, p.720-3, 1988) to compute packet error probabilities in frequency-hopped spread-spectrum packet radio networks with memoryless frequency-hopping patterns. The nature of the error is that the sequence of symbol errors is not Markov, so that a formula which is instrumental for the computation of packet error probabilities, is incorrect. He obtains exact and easy to compute expressions for the evaluation of the packet error probabilities.>

We consider the ART1 neural network architecture. It is shown that in the fast learning case, an ART1 network that is repeatedly presented with an arbitrary list of binary input patterns, self-stabilizes the recognition code of every size-l pattern in at most l list presentations.

A parallel hardware implementation of the associative memory neural network introduced by Hopfield is described. The design utilizes the Geometric Arithmetic Parallel Processor (GAPP), a commercially available single-chip VLSI general-purpose array processor consisting of 72 processing elements. The ability to cascade these chips allows large arrays of processors to be easily constructed and used to implement the Hopfield network. The memory requirements and processing times of such arrays are analyzed based on the number of nodes in the network and the number of exemplar patterns. Compared with other digital implementations, this design yields significant improvements in runtime performance and offers the capability of using large neural network associative memories in real-time applications.

The multiple-access capability of an asynchronous, frequency-hopped spread-spectrum (FH-SS) communication system employing error correcting codes is reviewed. Many current FH-SS systems employ error correcting codes with more than one code symbol per dwell interval. A method of computing the codeword error probability induced in such spread-spectrum systems is presented. The codeword error probabilities induced in spread-spectrum systems utilizing more than one versus one code symbol per dwell interval are compared. It is concluded that FH-SS systems with one code symbol per dwell interval are more efficient in combating multiple-access interference than FH-SS systems with more than one code symbol per dwell interval.>

The problem of computing packet error probabilities caused by a combination of noise at the receivers and interference between packet transmissions is discussed. An upper bound on the packet error probability induced in direct-sequence spread-spectrum networks is computed. The bound is valid independently of whether signals arrive with equal or unequal powers at the receiver site. It has a simple form, is easy to compute, and is valid for various classes of forward error correction codes (e.g. BCH, convolutional codes).>

The packet error probability induced in a frequency-hopped spread-spectrum packet radio network is computed. The frequency spectrum is divided into q frequency bins. Each packet is exactly one codeword from an (M, L) Reed-Solomon code (M=number of codeword symbols (bytes); L=number of information symbols (bytes)). Every user in the network sends each of the M bytes of his packet at a frequency chosen among the q frequencies with equal probability and independently of the frequencies chosen for other bytes (i.e., memoryless frequency-hopping patterns). Statistically independent frequency-hopping patterns correspond to different users in the network. Provided that K users have simultaneously transmitted their packets on the channel and a receiver has locked on to one of these K packets, the probability that this packet is not decoded correctly is evaluated. It is also shown that although memoryless frequency-hopping patterns are utilized, the byte errors at the receiver are not statistically independent; instead they exhibit a Markovian structure.>

In this correspondence, we consider the random accessing of a single slotted channel by a large number of packet transmitting users, whose cumulative traffic is Poisson. We assume the existence of the same feedback as that of the MCRAI protocols of Georgiadis, and full channel sensing, and we develop collision resolution algorithms that utilize the absorption concept of Gallager. We observe the improvement in the throughputs induced by the absorption, as well as the improved delay characteristics. Finally, we draw some conclusions about the limitations of the absorption idea.