Citation Context

...wn to compare favorably to other algorithms, (e.g., decision trees and rules), on a wide range of domains in which feature values were either numeric or binary (e.g., Aha, 1989; Aha and Kibler, 1989; =-=Aha et al., 1991-=-; Salzberg, 1989). This paper presents similar evidence in terms of classification accuracy for domains with symbolic feature values. However, before describing these domains, we should consider other...

Citation Context

...o the inherent difficulty of the domain or to the lack of sufficient data for learning. For example, instance-based learning programs (also called exemplar-based (Salzberg, 1990) or nearest neighbor (=-=Cover and Hart, 1967-=-) methods), which learn by storing examples as points in a feature space, require some means of measuring distance between examples (Aha, 1989; Aha and Kibler, 1989; Salzberg, 1989; Cost and Salzberg,...

Citation Context

...re sophisticated instance-based algorithm designed for domains in which some or all of the feature values are sym2 bolic. Our algorithm constructs modified "value difference" tables (in the =-=style of (Stanfill and Waltz, 1986)) to prod-=-uce a non-Euclidean distance metric, and we introduce the idea of "exception spaces" that result when weights are attached to individual examples. The combination of these two techniques res...

Citation Context

...back propagation can overcome this problem. Back propagation is a gradient descent method that propagates error signals back through a multi-layer network. It has been described in many places (e.g., =-=Rumelhart et al., 1986-=-; Rumelhart and McClelland, 1986), and readers interested in a more detailed description should look there. We will use the decision tree algorithm ID3 (Quinlan, 1986) as the basis for comparison with...

Citation Context

...ction of DNA promoter sequences. These domains have received considerable attention from connectionist researchers who employed the back propagation learning algorithm (Sejnowski and Rosenberg, 1986; =-=Qian and Sejnowski, 1988-=-; Towell et al., 1990). In addition, the word pronunciation problem has 3 been the subject of a number of comparisons using other machine learning algorithms (Stanfill and Waltz, 1986; Shavlik et al.,...

Citation Context

...o predict secondary structure involve the classification of residues into three categories: ff helix, fi sheet, and coil. Three of the most widely used approaches to this problem are those of Robson (=-=Garnier et al., 1978-=-), Chou and Fasman (1978), and Lim (1974), which produce classification accuracies ranging from 48% to 58%. Other, more accurate techniques have been developed for predicting tertiary from secondary s...

Citation Context

...ences. These domains have received considerable attention from connectionist researchers who employed the back propagation learning algorithm (Sejnowski and Rosenberg, 1986; Qian and Sejnowski, 1988; =-=Towell et al., 1990-=-). In addition, the word pronunciation problem has 3 been the subject of a number of comparisons using other machine learning algorithms (Stanfill and Waltz, 1986; Shavlik et al., 1989; Dietterich et ...

Citation Context

...back propagation can overcome this problem. Back propagation is a gradient descent method that propagates error signals back through a multi-layer network. It has been described in many places (e.g., =-=Rumelhart et al., 1986-=-; Rumelhart and McClelland, 1986), and readers interested in a more detailed description should look there. We will use the decision tree algorithm ID3 (Quinlan, 1986) as the basis for comparison with...

Citation Context

...r--Based Learning System. 2 For clarity, we use the term "example" to mean a training or test example being shown to the system for the first time. We use the term "exemplar" (foll=-=owing the usage of (Salzberg, 1991)) to refe-=-r specifically to an instance that has been previously stored in computer memory. Such exemplars may have additional information attached to them (e.g., weights). The term "instance " covers...

Citation Context

...Sejnowski, 1988; Towell et al., 1990). In addition, the word pronunciation problem has 3 been the subject of a number of comparisons using other machine learning algorithms (Stanfill and Waltz, 1986; =-=Shavlik et al., 1989-=-; Dietterich et al., 1990). All of these domains represent problems of considerable practical importance, and all have symbolic feature values, which makes them difficult for conventional nearest neig...

Citation Context

...erimental results show that a modified value-difference metric can process symbolic values exceptionally well. These results, taken together with other results on domains with numeric features (e.g., =-=Aha, 1990-=-; Aha et al., 1991), show that IBL algorithms perform quite well on a wide range of problems. 2 Learning algorithms Back propagation is the most widely used and understood neural net learning algorith...

Citation Context

...om the machine learning research community: the word pronunciation task (Sejnowski and Rosenberg, 1986; Shavlik et al., 1989), the prediction of protein secondary structure (Qian and Sejnowski, 1988; =-=Holley and Karplus, 1989-=-), and the prediction of DNA promoter sequences (Towell et al., 1989). Each domain has only symbolic-valued features; thus, our MVDM is applicable whereas standard Euclidean distance is not. Sections ...

Citation Context

...l et al., 1990). In addition, the word pronunciation problem has 3 been the subject of a number of comparisons using other machine learning algorithms (Stanfill and Waltz, 1986; Shavlik et al., 1989; =-=Dietterich et al., 1990-=-). All of these domains represent problems of considerable practical importance, and all have symbolic feature values, which makes them difficult for conventional nearest neighbor algorithms. We will ...

Citation Context

...ar-based (Salzberg, 1990) or nearest neighbor (Cover and Hart, 1967) methods), which learn by storing examples as points in a feature space, require some means of measuring distance between examples (=-=Aha, 1989-=-; Aha and Kibler, 1989; Salzberg, 1989; Cost and Salzberg, 1990). An example is usually a vector of feature values plus a category label. When the features are numeric, normalized Euclidean distance c...

Citation Context

.... Neural nets do not yet provide any insight into why they made the classification they did, although some recent efforts have explored new methods for understanding the content of a trained network (=-=Hanson and Burr, 1990-=-). It is also relatively easy to modify our algorithm to include domain specific knowledge: if the relative importance of the features is known, the features may be weighted accordingly in the distanc...

Citation Context

...equently used measure of performance in this domain are the correlation coefficients, which provide a measure of accuracy for each of the categories. They are defined by the following equation, from (=-=Mathews, 1975-=-): C ff = p ff \Theta n ff \Gamma u ff \Theta o ff q (n ff + u ff ) (n ff + o ff ) (p ff + u ff ) (p ff + o ff ) where p ff is the number times ff was correctly predicted, n ff is the number of times ...

Citation Context

...ception. The capability becomes even more important for IBL models that store only a subset of the training examples, because it further reduces 17 the number of points which must be stored (Cost and =-=Salzberg, 1990-=-). Given the above discussion, it should be clear that all instances should not be initialized with weights of 1. Consider a system trained on n \Gamma 1 instances, now training on the n th . A hierar...

Citation Context

...d by Towell et al. for several other machine learning algorithms, including back propagation, ID3, nearest neighbor with the overlap metric, and the best method reported in the biological literature (=-=O'Neill, 1989-=-). Recall that the overlap metric measures distance as the number of features with different values. It is also worth noting that in each of the 10 test runs of PEBLS, the same four instances caused t...

Citation Context

...an (1978), and Lim (1974), which produce classification accuracies ranging from 48% to 58%. Other, more accurate techniques have been developed for predicting tertiary from secondary structure (e.g., =-=Cohen et al., 1986-=-; Lathrop et al., 1987), but the accurate prediction of secondary structure has proven to be an extremely difficult task. The learning problem can be described as follows. A protein consists of a sequ...

Citation Context

...protein structure prediction, combining nearest neighbor with neural net learning and statistical information. Figures have not yet been published, but their method also outperforms previous methods (=-=Waltz, 1990-=-), although its accuracy does not exceed that of PEBLS. For the DNA promoter sequence prediction, Towell et al. (1990) report that KBANN, a technique that integrates neural nets and domain knowledge, ...

Citation Context

...1974), which produce classification accuracies ranging from 48% to 58%. Other, more accurate techniques have been developed for predicting tertiary from secondary structure (e.g., Cohen et al., 1986; =-=Lathrop et al., 1987-=-), but the accurate prediction of secondary structure has proven to be an extremely difficult task. The learning problem can be described as follows. A protein consists of a sequence of amino acids bo...

Citation Context

...assification time may be reduced to O(d log n). We implemented our system on a set of four loosely-coupled transputers as well as on a conventional architecture, and other recent efforts such as FGP (=-=Fertig and Gelernter, 1991-=-) use larger numbers of parallel processors. The MBRtalk system of Waltz and Stanfill (1986) implemented a form of k-nearest-neighbor learning using a tightly-coupled massively parallel architecture, ...

Citation Context

...est neighbor (Cover and Hart, 1967) methods), which learn by storing examples as points in a feature space, require some means of measuring distance between examples (Aha, 1989; Aha and Kibler, 1989; =-=Salzberg, 1989-=-; Cost and Salzberg, 1990). An example is usually a vector of feature values plus a category label. When the features are numeric, normalized Euclidean distance can be used to compare examples. Howeve...

Citation Context

...ver and Hart, 1967) methods), which learn by storing examples as points in a feature space, require some means of measuring distance between examples (Aha, 1989; Aha and Kibler, 1989; Salzberg, 1989; =-=Cost and Salzberg, 1990-=-). An example is usually a vector of feature values plus a category label. When the features are numeric, normalized Euclidean distance can be used to compare examples. However, when the feature value...

Citation Context

...alzberg, 1990) or nearest neighbor (Cover and Hart, 1967) methods), which learn by storing examples as points in a feature space, require some means of measuring distance between examples (Aha, 1989; =-=Aha and Kibler, 1989-=-; Salzberg, 1989; Cost and Salzberg, 1990). An example is usually a vector of feature values plus a category label. When the features are numeric, normalized Euclidean distance can be used to compare ...