Niklasson, L.F. & T. van Gelder, 1994. Can Connectionist Models Exhibit Non-Classical Structure Sensitivity, Proceedings of the Sixteenth Annual Conference of the Cognitive Science Society, 664--669.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....encountered before. There are a number of different definitions of structural systematicity. Three levels of systematicity from weak to strong were defined by (Hadley, 43 1992) but perhaps the most precise definition, using six levels of systematicity has been given by Niklasson and van Gelder (Niklasson van Gelder, 1994). These levels are: 1. Where no novelty is present in the inputs presented to the system. Every input presented to the system has already been encountered (for example in the training phase of a connectionist system) 2. Where novelty is present in the inputs, but all the individual arguments in ....

....representations are heavily dependent on the make up of the training set used to develop these representations, hence they often do not display systematicity beyond the second level described above. However, in performing simple logical operations with a system using distributed representations (Niklasson van Gelder, 1994) systematic performance of a neural network both when exposed to novel inputs and when exposed to inputs appearing in novel positions has been demonstrated (i.e. up to the fourth level above) 7. Conclusions We reviewed here a formalism that facilitates variable binding in connectionist networks. ....

NIKLASSON, L. and T. van GELDER. (1994) Can connectionist models exhibit non-classical structure sensitivity? Proceedings of the Cognitive Science Society, 664-669.

....to Fodor and Pylyshyn [1988] but in common with Hinton, McClelland and Rumelhart [1986] I do see implementation issues as interesting and important because implementation choices often determine which operations are fast and easy to compute and which are slow and difficult. Chalmers [1990] and Niklasson and van Gelder [1994] note that connectionist representations for compositional data structures are especially interesting because they raise the possibility of content sensitive manipulation without decomposition, which appears to have no analogue in conventional symbol based representations. Some authors use the ....

....like (LOVED X Y) to ones for (LOVED Y X) where the reduced descriptions were found by a RAAM. Chalmers [1990] trained a feedforward network to transform reduced descriptions for simple passive sentences to ones for active sentences, where again the reduced descriptions were found by a RAAM. Niklasson and van Gelder [1994] trained a feedforward network to do material conditional inference (and its reverse) on reduced descriptions found by a RAAM. This involves transforming reduced descriptions for formulae of the form (A B) to ones of the form ( A B) and vice versa) Legendre, Miyata, and Smolensky [1991] ....

Niklasson, L. and van Gelder, T. 1994. Can connectionist models exhibit non-classical structure sensitivity? In Proceedings of the Sixteenth Annual Conference of the The Cognitive Science Society.

.... by associative architectures (Fodor Pylyshyn, 1988) On the other hand it has often been shown that human reasoning does not readily conform to logic, and is better modelled by more content specific processes such as mental models, that function as analogues (Johnson Laird Byrne, 1991; Niklasson van Gelder, 1994). There has been a parallel debate about whether learning is conscious or unconscious, and whether what is learned is rule based or instance based (Shanks St. John, 1994) These dichotomies also overlap to some extent with the implicit explicit distinction (Clark Karmiloff Smith, 1993) As ....

Niklasson, L., & van Gelder, T. (1994, ). Can connectionist models exhibit non-classical structure sensitivity? Sixteenth Annual Conference of the Cognitive Science Society, Atlanta, Georgia, 664-669.

.... et al. 1986) and RAAM (Pollack, 1990) Compared to earlier systems, they had the advantage of compositionality built more explicitly into their design, and they have shown a great deal of promise in a number of areas (Chalmers, 1990, Elman, 1990, Chrisman, 1991, Blank et al. 1992, Plate, 1994, Niklasson van Gelder, 1994). However the data used to test these models has generally been confined to relatively simple structures at most 3 or 4 levels deep. Our aim was to see whether one of these architectures could be adapted to handle linguistic data of real world complexity. To find such data, we took a small ....

Niklasson, L.F. & T. van Gelder, 1994. Can Connectionist Models Exhibit Non-Classical Structure Sensitivity, Proceedings of the Sixteenth Annual Conference of the Cognitive Science Society, 664--669.

....like (LOVED X Y) to ones for (LOVED Y X) where the reduced descriptions were found by a RAAM. Chalmers [1990] trained a feedforward network to transform reduced descriptions for simple passive sentences to ones for active sentences, where again the reduced descriptions were found by a RAAM. Niklasson and van Gelder [1994] trained a feedforward network to do material conditional inference (and its reverse) on reduced descriptions found by a RAAM. This involves transforming reduced descriptions for formulae of the form (A B) to ones of the form ( A B) and vice versa) Legendre, Miyata, and Smolensky [1991] ....

Niklasson, L. and T. van Gelder (1994). Can connectionist models exhibit non-classical structure sensitivity? In Proceedings of the Sixteenth Annual Conference of The Cognitive Science Society.

....hidden patterns generated by a RAAM can be used as input (and output) in other networks. It has been shown that holistic, but systematic, operations can be performed on the hidden patterns without resorting to decoding them, in search for certain constituents, e.g. Pollack, 1990; Chalmers, 1990; Niklasson and van Gelder, 1994; Chrisman, 1991. We will in this paper: identify the actual reasons for the success of holistic operations on patterns generated by recursive connectionist systems, 1. A bottleneck network has less units in the hidden layer than the in input and output layer. verify that structured ....

....idea that holistic computation can be applied to the structure, encoded by the RAAM. The structure is preserved and can be used for generalisation. This feature can be used for different types of transformations, e.g. transformations to other structures (which is the case in Chalmers, 1991 and Niklasson and van Gelder 1994), or extraction of information about the tree structure, e.g. the depth and balance examples presented here. The take home message is if the task is to separate different expressions, use different tree structures to represent the different types. We have further shown that the holistic ....

Niklasson, L. F. and van Gelder T., (1994), Can Connectionist Models Exhibit Non-Classical Structure Sensitivity, Proceedings of the Cognitive Science Society, pp 664 - 669.

....approach for this classification appears to be to choose the token with the shortest Euclidean distance to the decoded representation. Several simulations have demonstrated that these techniques can be used for stable decomposition of this type of superpositional representations (Niklasson, 1993; Niklasson and van Gelder, 1994). It appears that RAAMs satisfy van Gelder s definition of a functional compositional scheme, since they possess techniques for both encoding and decoding 4 There is, however, an additional component of this issue which van Gelder was neutral towards; structured representations. In a PSS ....

Niklasson, L. and van Gelder, T. (1994). Can connectionist models exhibit non-classical structure sensitivity? In Proceedings of the Cognitive Science Society. Lawrence Erlbaum.

....approach for this classification appears to be to choose the token with the shortest Euclidean distance to the decoded representation. Several simulations have demonstrated that these techniques can be used for stable decomposition of this type of superpositional representations (Niklasson 1993; Niklasson van Gelder 1994). It appears that RAAMs satisfy van Gelder s definition of a functional compositional scheme, since they possess techniques for both encoding and decoding. 4 There is, however, an additional component of this issue which van Gelder was neutral towards; structured representations. In a PSS ....

Niklasson, L. and van Gelder, T. (1994). Can connectionist models exhibit non-classical structure sensitivity? In Proceedings of the Cognitive Science Society, pages 664--669. Lawrence Erlbaum.

No context found.

Niklasson, L.F. & T. van Gelder, 1994. Can Connectionist Models Exhibit Non-Classical Structure Sensitivity, Proceedings of the Sixteenth Annual Conference of the Cognitive Science Society, 664--669.

No context found.

L. Niklasson and T. van Gelder, \Can connectionist models exhibit non-classical structure sensitivity?," in Proceedings of the 16

No context found.

Niklasson, L., & van Gelder, T. (1994). Can connectionist models exhibit nonclassical structure sensitivity. In A. Ram & K. Eiselt (Eds.), Proceedings of the Sixteenth Annual ConferenceoftheCognitive ScienceSociety, pp. 644-- 669. Lawrence Erlbaum.

No context found.

Niklasson, L. and T. van Gelder (1994). Can connectionist models exhibit non-classical structure sensitivity? In Proceedings of the Sixteenth Annual Conference of The Cognitive Science Society.

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