Movellan, J. R., and McClelland, J. L. 1992. Learning continuous probability distributions with symmetric diffusion networks. Cognitive Science 17, 463-496.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....w ji . The derivation here was for the simplest sort of Boltzmann Machine, with binary units and only pairwise connections between the units. However, the technique is immediately applicable to higher order Boltzmann Machines (Hinton, 1987) as well as to Boltzmann Machines with non binary units (Movellan and McClelland, 1991). 4.4 Weight Perturbation In weight perturbation (Jabri and Flower, 1991; Alspector et al. 1993; Flower and Jabri, 1993; Kirk et al. 1993; Cauwenberghs, 1993) the gradient r w is approximated using only the globally broadcast result of the computation of E(w) This is done by adding a random ....

Movellan, J. R. and McClelland, J. L. (1991). Learning continuous probability distributions with the contrastive Hebbian algorithm. Technical Report PDP.CNS.91.2, Carnegie Mellon University Dept. of Psychology, Pittsburgh, PA.

....is not particularly biologically plausible, we consider it to be an e#cient search technique for #nding networks that perform the task. More biologically plausible learning schemes exist, but they tend to be extremely slow to converge, while often achieving basically the same end result #Movellan and McClelland, 1993; Plaut, 1991#. In this chapter, we #rst consider what some of the possible dimensions of #feature space are, and illustrate these with a discussion of the features used in some recent engineeringoriented face recognition models. In doing so, we hope to clarify to some extent what kinds of ....

Movellan, J. R. and McClelland, J. L. #1993#. Learning continuous probability distributions with symmetric di#usion networks. Cognitive Science, 17#4#:463#496.

....algorithms and fuzzy logic. By these methods it is possible to use even more information than before, because now it does not have to be certain information encoded in terms of right or wrong, on or off, or zero or one. One problem that I found unsettling about neural networks (Johansson, 1995a; Movellan McClelland, 1993) was that the method had severe problems with the shift of perspective. Whenever two alternatives were present, back propagation networks tried to make a compromise and blend the different alternatives into a definite non solution. In (Johansson, 1995a) it was shown that the possibility of having ....

....a compromise and blend the different alternatives into a definite non solution. In (Johansson, 1995a) it was shown that the possibility of having two correct past tense forms in Swedish lead to an impossible task for the neural net to learn, unless different underlying forms are constructed. (Movellan McClelland, 1993) used the example of training a car auto pilot on turning left and turning right, and to test the program with a fork in the road. They also showed that there is a possibility to activate solutions at different times by using a SDN stochastic diffusion network . To be able to resolve choice ....

Movellan, J.R., & McClelland, J.L. (1993). Learning Continuous Probability Distributions with Symmetric Diffusion Networks. Cognitive Science, 17, 463-496.

....j w ji . The derivation here was for the simplest sort of Boltzmann Machine, with binary units and only pairwise connections between the units. However, the technique is immediately applicable to higher order Boltzmann Machines (Hinton, 1987) as well as to Boltzmann Machines with non binary units (Movellan and McClelland, 1991). 4.4 Weight Perturbation In weight perturbation (Jabri and Flower, 1991; Alspector et al. 1993; Flower and Jabri, 1993; Kirk et al. 1993; Cauwenberghs, 1993) the gradient r w is approximated using only the globally broadcast result of the computation of E(w) This is done by adding a random ....

Movellan, J. R. and McClelland, J. L. (1991). Learning continuous probability distributions with the contrastive Hebbian algorithm. Technical Report PDP.CNS.91.2, Carnegie Mellon University Dept. of Psychology, Pittsburgh, PA.

....an entire mapping from inputs into probability densities on R d . Note this is a much harder problem than regression, since we care about the entire density of outputs, not just its expected value. In general density estimation is important when unimodal uncertainty models are not appropriate (Movellan McClelland, 1993). A popular cost function for density estimation is the Kullback Leibler information criterion (KLIC) In our case we need the KLIC between the desired and obtained conditional distributions averaged w.r.t the input density, Haykin, 1994, p. 447) C( Z R m dx fX (x) Z R d do p(ojx) ....

.... Boltzmann learning algorithm: C( i = fi Z R m dxfX (x) Cov (R ; S i j X = x) 58) Gammafi Z R m dx fX (x) Z R d do p(ojx) E ( S i j O = o; X = x ) Gamma E ( S i jX = x ) For applications of diffusion networks to density estimation problems, see Movellan and McClelland (1993). Methods of density estimation are discussed in Stark and McClelland (1994) Risk Minimization The objective in this case is to minimize the expected loss, which in Bayesian decision theory is known as the risk (Duda Hart, 1973, p. 14) C( E [ ae(O( X( 59) Risk minimization is ....

Movellan, J. R. & McClelland, J. L. (1993). Learning Continuous Probability Distributions with Symmetric Diffusion Networks. Cognitive Science, 17, 463--496.

No context found.

Movellan, J. R., and McClelland, J. L. 1992. Learning continuous probability distributions with symmetric diffusion networks. Cognitive Science 17, 463-496.

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