N. Intrator. A neural network for feature extraction. In D. Touretzky, editor, Neural Information Processing Systems, volume 2, pages 719 726. Morgan Kaufmann, San Mateo, CA, 1990.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... for feature extraction is projection pursuit (for a review, see [9] The idea behind projection pursuit is to pick interesting low dimensional projections of a high dimensional point cloud, by maximizing an objective function such as the deviation of the projected distribution from normality [10]. In the present work we chose to explore a considerably simpler method of feature extraction, based on the notion of localized receptive field, borrowed from neurobiology of vision. The receptive field (RF) of a neuron anywhere in the visual pathway is defined as that portion of the retinal ....

N. Intrator. A neural network for feature extraction. In D. Touretzky, editor, Neural Information Processing Systems, volume 2, pages 719 726. Morgan Kaufmann, San Mateo, CA, 1990.

....first projective approximation estimator. Chen [3] showed that under appropriate conditions the rate of convergence of projection pursuit regression is independent of dimensionality d. Projection pursuit was first introduced in the context of learning networks by Barron and Barron [1] Intrator [23, 24, 25, 26] applied exploratory projection pursuit to do feature extraction in speech recognition. Maechler et al. [31] and Hwang et al. [17, 18, 19, 21, 22] studied several two dimensional examples to compare projection pursuit regression and one hidden layer sigmoidal neural networks. Moody and Yarvin [34] ....

Intrator, N., "A Neural Network for Feature Extraction". in D. S. Touretzky (ed.) Advances in Neural Information Processing Systems 2. San Mateo, CA: Morgan Kaufmann, 1990.

....(Zetzsche, 1997) that for natural images, random local projections yield somewhat longer tailed distributions than Gaussian. We can still justify this approach, because interesting structure can still be found in non random directions which yield projections that are farther from Gaussian. Intrator (1990) has shown that a BCM neuron can find structure in the input distribution that exhibits deviation from Gaussian distribution in the form of multi modality in the projected distributions. This type of deviation, which is measured by the first three moments of the distribution, is particularly ....

To appear. Intrator, N. (1990). A neural network for feature extraction. In Touretzky, D. S. and Lippmann, R. P., editors, Advances in Neural Information Processing Systems, volume 2, pages 719--726. Morgan Kaufmann, San Mateo, CA.

....for extending the algorithm by incorporating additional information in the estimator. We share the goal of seeking highly informative, bimodal projections of the input with the Bienenstock Cooper Munro (BCM) algorithm (Bienenstock et al. 1982) In our notation, the BCM learning rule according to (Intrator, 1990, 1991a, 1991b, 1992) is Delta w f 0 (y) x z 2 Gamma 4 3 z D z 2 E : III:17) Whereas in the derivation of BINGO some of the nonlinearities cancel (III.11) yielding a relatively straightforward learning rule, here they compound into a rather complicated polynomial. The ....

Intrator, N. (1990). A neural network for feature extraction. In (Touretzky, 1990), pages 719--726.

....(Zetzsche, 1997) that for natural images, random local projections yield somewhat longer tailed distributions than Gaussian. We can still justify this approach, because interesting structure can still be found in non random directions which yield projections that are farther from Gaussian. Intrator (1990) has shown that a BCM neuron can find structure in the input distribution that exhibits deviation from Gaussian distribution in the form of multi modality in the projected distributions. This type of deviation, which is measured by the first three moments of the distribution, is particularly ....

Intrator, N. (1990). A neural network for feature extraction. In Touretzky, D. S. and Lippmann, R. P., editors, Advances in Neural Information Processing Systems, volume 2, pages 719--726. Morgan Kaufmann, San Mateo, CA.

....Freedman (1984) show that for most high dimensional clouds (of points) most low dimensional projections are approximately Gaussian. This finding suggests that important information in the data is conveyed in those directions whose single dimensional projected distribution is far from Gaussian. Intrator (1990) has shown that a BCM neuron can find structure in the input distribution that exhibits deviation from Gaussian distribution in the form of multi modality in the projected distributions. Since clusters can not be found directly in the data due to its sparsity, this type of deviation, which is ....

Intrator, N. (1990). A neural network for feature extraction. In Touretzky, D. S. and Lippmann, R. P., editors, Advances in Neural Information Processing Systems, volume 2, pages 719--726. Morgan Kaufmann, San Mateo, CA.

.... in very high dimensional space suffers from the inherent sparsity of such space, and therefore can not be directly approached by classical methods such as cluster analysis (Duda and Hart, 1973) discriminant analysis (Fisher, 1936; Sebestyen, 1962) or factor analysis (Harman, 1967) Recent work (Intrator, 1990; Intrator and Cooper, 1992) connecting biologically motivated feature extraction networks (Bienenstock et al. 1982, henceforth to be referred to as BCM ) with sophisticated statistical techniques (Friedman and Tukey, 1974; Friedman, 1987) suggests that this 1 Edelman and Weinshall (1991) used ....

....The idea behind projection pursuit is to pick interesting low dimensional projections of a high dimensional point cloud by maximizing an objective function called the projection index. The projection index usually measures some form of deviation from normality of the projected distribution. 4 Intrator (1990) presented a multiple feature extraction method that seeks multi modality in the projected distributions. This method is based on a modified version of the BCM neuron (Bienenstock et al. 1982) The biological relevance of this neuron has been extensively studied (Bear et al. 1987; Bear and ....

[Article contains additional citation context not shown here]

Intrator, N. (1990). A neural network for feature extraction. In Touretzky, D. S. and Lippmann, R. P., editors, Advances in Neural Information Processing Systems, volume 2, pages 719--726. Morgan Kaufmann, San Mateo, CA.

.... In addition to a standard single hidden layer network trained using the back propagation algorithm, we applied a hybrid method that combines unsupervised feature extraction and classification [7] The unsupervised feature extraction is based on the biologically motived BCM neuron [3] and was shown [6] to be related to a statistical technique called exploratory projection pursuit [4] The hybrid method incorporates a single hidden layer ANN trained to minimize MSE using error back propagation in addition to minimizing a projection pursuit index [8] that favors multimodality in the projected ....

N. Intrator. A neural network for feature extraction. In D. S. Touretzky and R. P. Lippmann, editors, Advances in Neural Information Processing Systems, volume 2, pages 719--726. Morgan Kaufmann, San Mateo, CA, 1990.

....third order correlation C(ff; fi; fl) of the input activity to be C(ff; fi; fl) E[a(ff)a(fi)a(fl) yields dm(ff; t) dt = f X fi;fl C(ff; fi; fl)m(fi)m(fl) Gamma Theta m X fi m(fi)C(ff; fi)g: 5. 6) Using a definition for the threshold Theta m def = E[ P ff a(ff)m(ff) 2 (Intrator, 1990; Intrator and Cooper, 1992) and using the second order correlation function C, Theta m becomes Theta m = P fl ;ffi C(fl; ffi )m(fl)m(ffi) Therefore, eq. 5.6) becomes dm(ff; t) dt = f X fi;fl C(ff; fi; fl)m(fi)m(fl) Gamma X fifl ffi C(ff; fi)C(fl; ffi)m(fi)m(fl)m(ffi)g: 5.7) A ....

....dimensional projected distribution is far from Gaussian. For example, some known measures of deviation from normality are skewness and kurtosis which are functions of the first four moments of the distribution. These moments contain information about statistical correlations up to fourth order. Intrator (1990) has shown that a BCM neuron (given by equation 4.8) can find structure in the input distribution that exhibits deviation from normality in the form of multi modality in the projected distributions. This type of deviation, which is measured by the first three moments of the distribution, is ....

Intrator, N. (1990). A neural network for feature extraction. In Touretzky, D. S. and Lippmann, R. P., editors, Advances in Neural Information Processing Systems, volume 2, pages 719--726. Morgan Kaufmann, San Mateo, CA.

....singledimensional projected distribution is far from Gaussian. Polynomial moments are good candidates for measuring deviation from Gaussian distribution; for example, skewness and kurtosis which are functions of the first four moments of the distribution, are frequently used in this connection. Intrator (1990) has shown that a BCM 9 neuron can find structure in the data that exhibits deviation from normality in the form of multi modality in the projected distributions. Because clusters cannot be found directly in the data due to its sparsity (recall the curse of dimensionality) this type of ....

Intrator, N. (1990). A neural network for feature extraction. In Touretzky, D. S. and Lippmann, R. P., editors, Advances in Neural Information Processing Systems, volume 2, pages 719--726.

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