Niklasson, L. & Boden, M. (1997), Representing structure and structured representations in connectionist networks, in A. Browne, ed., `Neural Network Perspectives on Cognition and Adaptive Robotics', IOP Press, pp. 20--50.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....and AB (right) respectively. SOM 2 in figure b has one peak centered around winner for sequence Delta Delta, where we considered Delta = D = 1; 1] for convenience. For structure representations it is important that they themselves must be structured to allow structure sensitive processing [14]. To check this property of representations with SOMs, we have generated all 2 symbol sequences having either of the templates f(ffiffi)ffi; ffi(ffiffi) ffiffi) ffiffi)g. Each structure was represented by 7 Theta 7 SOM 1 and SOM 2 as Y = Y 1 ; Y 2 ] SOM 1 had only one input (because there are ....

L. Niklasson and M. Bod'en. Representing structure and structured representations in connectionist networks. In A. Browne, editor, Neural Network Perspectives on Cognition and Adaptive Robotics, pages 20--50. IOP Press, 1997.

....CRG is mainly focused on ANNs, also called Connectionist Networks or Parallel Distributed Processing. CRG was established in 1991 and has been conducting basic research on the implications of an ANN approach to cognition (Boden Narayanan, 1993, 1994; Boden 1994, 1996; Boden Niklasson, 1995; Niklasson Boden, 1997; Niklasson, 1993, 1996; Niklasson Sharkey, 1992, 1994, 1997; Niklasson van Gelder, 1994a, 1994b) and adaptive robotics (Ziemke, 1996a, 1996b, 1996c, 1996d, 1997a, 1997b, strm Ziemke, 1997, Sharkey Ziemke, 1999) In addition to this CRG members have also applied ANN techniques to real ....

.... techniques to real world problems such as radar signal processing, in particular oil spill detection (in cooperation with Ericsson Microwave Systems, Mlndal) and sales forecasting (ICA Handlarnas AB, Solna) see e.g. Ziemke, 1995, 1996e, 1996f; Ziemke Athley, 1995a, 1995b, Ziemke, Boden Niklasson, 1997). CRG has, in addition to the projects on oil spill detection and sales forecasting, been involved in some minor projects involving practical applications of ANNs; fish stock prediction (HS IDA EA 96 102) classification of logs for automatic separation in a saw mill (HS IDA EA 96 103) and ....

Niklasson, L. F. and Bodn, M. B., (1997) Representing Structure and Structured Representations in Connectionist Networks, in: Browne, A.J. (ed.) Neural Network Perspectives on Cognition and Adaptive Robotics, IOP Publishing, Bristol, UK.

....into B and A. RAAM generates representations which are not only capable of representing structure but also qualify as being structured although not in a symbolic sense with syntactic parts. The principal argument relies on some experiments on RAAM generated representations for tree structures (Niklasson and Bod en, 1997). The experiments showed that structural properties (e.g. tree depth) were detectable holistically by a simple classification network for which the representations were presented. 3.2 RAAM, representational space and content As has been hinted earlier, for a layer of receiving units (all ....

Niklasson, L. and Bod'en, M. (1997). On representing structure and structured representations in connectionist networks. In Browne, A., editor, Current Perspectives in Neural Computing. IOP.

....into B and A. RAAM generates representations which are not only capable of representing structure but also qualify as being structured although not in a symbolic sense with syntactic parts. The principal argument relies on some experiments on RAAM generated representations for tree structures (Niklasson and Bod en, 1997). These experiments showed that structural properties (e.g. tree depth) were detectable holistically by a simple classi cation network for which the representations were presented. 3.2 RAAM, representational space and content As has been hinted earlier, for a layer of receiving units (all ....

Niklasson, L. and Boden, M. (1997). On representing structure and structured representations in connectionist networks. In Browne, A., editor, Current Perspectives in Neural Computing, pages 20-50. IOP.

....into B and A. RAAM generates representations which are not only capable of representing structure but also qualify as being structured although not in a symbolic sense with syntactic parts. The principal argument relies on some experiments on RAAM generated representations for tree structures (Niklasson and Bod en, 1997). These experiments showed that structural properties (e.g. tree depth) were detectable holistically by a simple classification network for which the representations were presented. 3.2 RAAM, representational space and content As has been hinted earlier, for a layer of receiving units (all ....

Niklasson, L. and Bod'en, M. (1997). On representing structure and structured representations in connectionist networks. In Browne, A., editor, Current Perspectives in Neural Computing, pages 20--50. IOP.

....end up in a similar region in the representational space. 12 See for instance the differences in results between applying a hierarchical cluster technique on hidden representations compared to the analysis possible when training a separate network to detect specific features in them (Niklasson and Bod en, 1997). Burr, 1990; Sharkey and Jackson, 1995) If the receiving units are sigmoidal units, the regions are not as clear as when threshold units are used, but they do, nevertheless, extract content. Figure 2 shows how the activation space of two sending units (input x and input y) is divided by the ....

....processing elements. The classicalist representing world could be a parenthesized representation or a graph structure, both of which preserve the constituent structure of the expression. and characteristic features present in the training set for the training task (Bod en and Niklasson, 1995; Niklasson and Bod en, 1997). The output weights need to distinguish between hidden activities in terms of the constituent structure imposed by the particular input target presentation. Thus, the generated representations can be evaluated in terms of that constituency. For example, if learning is successful the pattern for ....

[Article contains additional citation context not shown here]

Niklasson, L. and Bod'en, M. (1997). On representing structure and structured representations in connectionist networks. In Browne, A., editor, Current Perspectives in Neural Computing. IOP.

No context found.

Niklasson, L. & Boden, M. (1997), Representing structure and structured representations in connectionist networks, in A. Browne, ed., `Neural Network Perspectives on Cognition and Adaptive Robotics', IOP Press, pp. 20--50.

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