Auer, P. (1997) On learning from mult-instance examples: Empirical evaluation of a theoretical approach. Proceedings 14th International Conference on Machine Learning, 160.166.1a-S (DD) 80.166.1a-S (DD) 160.166.1a-S (EM-DD) 80.166.1a-S (EM-DD)  Home/Search   Document Not in Database   Summary   Related Articles   Check

....paper [4] in which they developed MI algorithms for learning axis parallel rectangles (APRs) and they also provided two benchmark Musk data sets. Following this work, there has been a significant amount of research directed towards the development of MI algorithms using different learning models [2,5,6,9,12]. Maron and Raton [7] applied the multiple instance model to the task of recognizing a person from a series of images that are labeled positive if they contain the person and negative otherwise. The same technique was used to learn descriptions of natural scene images (such as a waterfall) and to ....

....from each positive bag and no instances from any negative bag. The resulting box was then expanded (via a statistical technique) to get better results. However, the test data from Musk1 was used to tune the parameters of the algorithm. These parameters are then used for Musk1 and Musk2. Auer [2] presented an algorithm, MULTINST, that learns using simple statistics to find the halfspaces defining the boundaries of the target APR and hence avoids some potentially hard computational problems that were required by the heuristics used in the iterated discrim algorithm. More recently, Wang and ....

[Article contains additional citation context not shown here]

Auer, P. (1997) On learning from mult-instance examples: Empirical evaluation of a theoretical approach. Proceedings 14th International Conference on Machine Learning, 160.166.1a-S (DD) 80.166.1a-S (DD) 160.166.1a-S (EM-DD) 80.166.1a-S (EM-DD)

....across 10 runs using 10 fold cross validation. Algorithm Musk1 Musk 2 accuracy accuracy EM DD 96.8 96.0 Iterated discrim [4] 92.4 89.2 Citation kNN [12] 92.4 86.3 Bayesian kNN [12] 90.2 82.4 Diverse density [6] 88.9 82.5 Multi instance neural network [9] 88.0 82.0 Multinst [2] 76.7 84.0 and produces a real valued output. For the musk data sets we simply use 1.0 (respectively, 0.0) as the label for positive (respectively, negative) bags. The boolean prediction is obtained by predicting positive i the predicted label is at least 0.5. We present results using one real ....

....positive bag and no examples from any negative bag. Then they expand the resulting box (via a statistical technique) to get better results. They use a 10 fold cross validation using the test data from Musk1 to appropriately tune their parameters. These parameters are then used for Musk2. Auer [2] presented an algorithm, Multinst, that learns using simple statistics to nd the halfspaces de ning the boundaries of the target APR and hence avoids some potentially hard computational problems that were required by the heuristics used in the iterated discrim algorithm. More recently, Wang and ....

Auer, P. (1997) On learning from mult-instance examples: Empirical evaluation of a theoretical approach. Proceedings 14th International Conference on Machine Learning, pp. 21-29. San Francisco, CA: Morgan Kaufmann.

.... each instance is a possible con guration (or shape) for a molecule of interest and each bag (example) contains all low energy (and hence likely) con gurations for the molecule (Dietterich, Lathrop, Lozano P erez, 1997) There has been a signi cant amount of research directed towards this problem (Auer 1997; Maron Lozano P erez, 1998, Maron 1998, Wang and Zucker 2000) Other applications for the multiple instance model have also been proposed (Maron Raton, 1998; Ru o 2000) Prior research performed under the multiple instance model is for concept learning (i.e. boolean labels) Binding anity ....

....of Jain et al. 1994) in which they presented COMPASS which as an APR like neural network algorithm which is robust to errors in the initial alignment of the molecules. While COMPASS can handle real valued labels, we are not aware of any reported results on any available real valued data sets. Auer (1997) presented an algorithm that learns using simple statistics and hence avoids some potentially hard computational problems that were required by the heuristics used by Dietterich et al. Their algorithm worked quite well on the Musk2 data set (obtaining a 84 accuracy) despite the fact that they ....

[Article contains additional citation context not shown here]

Auer, P. (1997) On learning from mult-instance examples: Empirical evaluation of a theoretical approach.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC