G. Carpenter and T. A., \Rule extraction: From neural architecture to symbolic representation," Connection Science, vol. 7, no. 1, pp. 3-27, 1995.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....network with descriptions of the domain knowledge of applied area. So that not only predictions can be made but also the reasons for the predictions can be explained. DNN to conventional neural networks is like econometrics to regression. It incorporates uncovered rules such as propositional rules [3, 9] and fuzzy rules [5] as well as domain knowledge to traditional neural networks. We expect that DNN networks could make more accurate and explainable predictions. Some issues in construction of DNN include knowledge based management, architecture enhancement, rule measurement criteria, threshold ....

G. A. Carpenter & A.W. Tan, "Rule Extraction: From Neural Architecture to Symbolic Representation", Connection Science, 7(1), pp3-27, 1995.

....in a cluster by ARTa and the whole character belongs to a fuzzy set with label l, with a certain membership function value. However, extraction of a reduced set of rules is difficult when the number of clusters becomes high, because of the supervision method employed in FasArt or Fuzzy ARTMAP [10]. In order to reduce network complexity, exploit the classifier generalization capabilities and simplify the process of rule fusion [10] to create allographs, reconstruction of the prototype from FasArt c weights is made, and unsupervised classification of these prototypes is carded out with an ....

.... extraction of a reduced set of rules is difficult when the number of clusters becomes high, because of the supervision method employed in FasArt or Fuzzy ARTMAP [10] In order to reduce network complexity, exploit the classifier generalization capabilities and simplify the process of rule fusion [10] to create allographs, reconstruction of the prototype from FasArt c weights is made, and unsupervised classification of these prototypes is carded out with an unsupervised module of FasArt, leading to a great reduction in the number of clusters. A typical example of this methodology is the ....

G.A. Carpenter and H.A. Tan, "Rule extraction: From neural architecture to symbolic representation", Connection Science, 7(1):3-27, 1995.

....advantages of adaptive resonance theory (ART) networks. In Fuzzy ARTMAP each category in the field (Fig. 1) roughly corresponds to a rule. Each node is defined by a weight vector that can be directly translated into a verbal or algorithmic description of the antecedents of the corresponding rule [7]. Though Fuzzy ARTMAP inherently represents acquired knowledge in the form of IF THEN rules, large or noisy Manuscript received December 13, 2000; revised April 12, 2001. This work was supported in part by Spanish CICYT under Project TIC1999 0446 C02 01. E. Gmez Snchez and Y. A. Dimitriadis are ....

....J. M. Cano Izquierdo and J. Lpez Coronado are with the Department of System Engineering and Automatic Control, Polytechnical University of Cartagena, Murcia, Spain. Publisher Item Identifier S 1045 9227(02)00347 8. datasets typically cause Fuzzy ARTMAP to generate too many rules [7]. This problem is known as category proliferation [8] It is due to the application of the match tracking mechanism, that however is necessary to guarantee fast, accurate, on line learning. This mechanism is fired after a pattern has been presented, if the selected category in ART predicts a wrong ....

[Article contains additional citation context not shown here]

A. Carpenter and H. A. Tan, "Rule extraction: From neural architecture to symbolic representation," Connection Sci., vol. 7, no. 1, pp. 3--27, 1995.

....will emerge. In immune signalling networks, both natural and artificial, it is probable that the emergent connection structure will be difficult to explain in detail. This has proved to be the case in the relatively simple problems of constructing neural networks to carry out defined tasks [7] [4]. Nonetheless we believe that considerations such as those suggested here can yield valuable overall understanding of a system whose detailed workings may forever be veiled in mystery. 9 A brief comparison to some other approaches to decentralized systems What can the immune system teach us ....

G.A. Carpenter and A-H. Tan, "Rule extraction: From neural architecture to symbolic representation", Connections Science, Vol. 7, 1995, pp. 3--28.

....pattern classification problems due to their ability to model nonlinear relationship. Neural networks are also robust in handling data with noise or missing values due to their inherently parallel data processing. Recent development in algorithms which extract rules from trained neural networks [1, 2, 7, 15, 16, 17, 21, 22] can be expected to lead to an even wider use of neural networks for pattern classification as the extracted rules add a very useful feature to the network classification process. Many of these algorithms extract classification rules that preserve the high accuracy of the network from which they ....

G.A. Carpenter and A-H. Tan, Rule extraction: From neural architecture to symbolic representation, Connection Science (7)1 (1995) 3--28.

....extent to which the underlying neural network incorporates specialised training regimes) The expressive power of the rules describes the format of the extracted rules. Currently there exist rule knowledge extraction techniques that extract rules in various formats including propositional rules [9][10[11] fuzzy rules [12] 13] scientific laws [14] finite state automata [15] decision trees [16] and m of n rules [17] The rule quality criterion is assessed via four characteristics, viz. a) rule accuracy, the extent to which the rule set is able to classify a set of previously unseen ....

G.A. Carpenter & A.W. Tan, Rule Extraction: From Neural Architecture to Symbolic Representation, Connection Science Vol 7 No 1, (1995), 3-27.

....and learning capabilities of the fuzzy hart s, and also suggested that the developed clusterings are not unrelated to those produced by a classical and a machine learning clustering algorithm. This allows us to address issues such as extracting rules from trained fuzzy hart s networks (see [7], for example) and or optimising the developed hierarchies. Along the latter line, it may be useful to view fuzzy hart s (especially in fast learning mode) as implementing the first sorting stage in a multi stage clustering mechanism [9] Furthermore, the component Fuzzy ART networks have ....

G.A. Carpenter and Ah-Hwee Tan. Rule extraction: from neural architecture to symbolic representation. Connection Science, 7(1):3--27, 1995.

....and verified by human experts. They can also provide new insights into the application problems and the corresponding data. It is not surprising that in recent years there has been a significant amount of work devoted to the development of algorithms that extract rules from neural networks [1, 4, 7, 8, 18, 19, 21, 23]. In order to obtain a concise set of symbolic rules, redundant and irrelevant units and connections of a trained neural network are usually removed by a network pruning algorithm before rules are extracted [3, 16, 18, 19] This process can be time consuming as most algorithms for neural network ....

G.A. Carpenter and A-H. Tan, "Rule extraction: From neural architecture to symbolic representation," Connection Science, vol. 7, no. 1, pp. 3--28, 1995.

....to be closely interrelated. For example, many connectionist algorithms have been proposed eliminating weights or units from networks in order to speed up learning, to avoid overfitting, and to produce interpretable networks (Mozer Smolensky, 1989; Le Cun et al. 1990; Hassibi et al. 1992; Carpenter Tan, 1995). bp som addresses these issues from a different angle: it uses a pruning criterion, but it also implements an active learning process (i.e. the influence of the soms) that stimulates certain hidden units to be inactive. The issue of overfitting is well known in the field of symbolic machine ....

Carpenter, G. A. & Tan, A.-H. (1995). Rule extraction: from neural architecture to symbolic representation. Connection Science, 7, 3--27.

.... categorization of complex databases (e.g. Asfour, Carpenter, and Grossberg, 1995; Asfour et al. 1993; Bachelder, Waxman, and Seibert, 1993; Baloch and Waxman, 1991; Bradski and Grossberg, 1995; Carpenter et al. 1992; Carpenter, Grossberg, and Reynolds, 1991, 1995; Carpenter and Ross, 1994; Carpenter and Tan, 1995; Caudell, Smith, Escobedo, and Anderson, 1994; Dubrawski and Crowley, 1994; Gjerdingen, 1990; Goodman et al. 1992; Ham and Han, 1993; Harvey, 1993; Kasperkiewicz, Racz, and Dubrawski, 1994; Keyvan, Durg, and Rabelo, 1993; Metha, Vij, and Rabelo, 1993; Moya, Koch, and Hostetler, 1993; Seibert and ....

Carpenter, G.A. and Tan, A.-H. (1995). Rule extraction: From neural architecture to symbolic representation. Connection Science, 7, 3--27.

.... the recognition categories learned by the F 2 category nodes are compatible with rules that link antecedents to consequents (Figure 1) At any point during the incremental learning process, the system architecture can be translated into a compact set of rules analyzable by human experts [6]. Examples Rule Insertion Refined Rules Original Rules Cascade ARTMAP Refined Cascade ARTMAP Extraction Rule Rule Refinement Figure 2: Cascade ARTMAP for symbolic knowledge refinementandevaluation. On the other hand, rules can be readily inserted into an ARTMAP network that can then ....

G. A. Carpenter and A.-H. Tan. Rule extraction: From neural architecture to symbolic representation. Connection Science, 7(1):3--27, 1995.

....networks. Also, by the self stabilizing property, learning in Cascade ARTMAP does not wash away existing knowledge and the meanings of units do not shift. This allows preservation of symbolic rules during neural network learning. Using a generalized ARTMAP rule extraction procedure [6] [7], the final system states can be converted back to a compact set of rules. This enables direct comparison of the original knowledge with the refined rules. Also, each extracted rule is associated with a confidence factor that indicates its importance or usefulness. This allows ranking and ....

....consequents. After match tracking, a new prediction loop then repeats from stage 2. F. Rule Extraction As a direct generalization of fuzzy ARTMAP, Cascade ARTMAP architecture can be readily translated into a set of symbolic rules using a generalized ARTMAP rule extraction procedure [6] [7]. A rule pruning procedure selects a small set of rules from Cascade ARTMAP networks based on their confidence factors. To derive concise rules, an antecedent pruning procedure aims to removeantecedents from rules while preserving accuracy. F.1 Rule Pruning The rule pruning algorithm derives a ....

G. A. Carpenter and A. H. Tan. Rule extraction: From neural architecture to symbolic representation. Connection Science, 7(1):3--27, 1995.

....each node a confidence factor computed in real time. A node with little or no reinforcement will have low confidence and can be removed once the confidence falls belowachosen threshold. The use of confidence factors has been found to be effective in pruning category nodes of fuzzy ARTMAP systems (Carpenter Tan, 1993, 1995# Tan, 1994a) 10 4 Concept Hierarchy Network 4.1 The Concept Hierarchy Representation In Touretzky (1986) is a and is not a links are used to connect various object classes as follows: elephant is a gray thing royal elephant is not a gray thing royal elephant is a elephant Figure 4: The ....

Carpenter, G. A., & Tan, A. H. (1995). Rule extraction: From neural architecture to symbolic representation. Connection Science, 7 (1), 3--27.

.... the recognition categories learned by the F a 2 category nodes are compatible with rules that link antecedents to consequents (Figure 1) At any point during the incremental learning process, the system architecture can be translated into a compact set of rules analyzable by human experts [6]. On the other hand, rules can be readily inserted into an ARTMAP network that can then be refined by learning from examples. To Examples Rule Insertion Refined Rules Original Rules Cascade ARTMAP Refined Cascade ARTMAP Extraction Rule Rule Refinement Figure 2: Cascade ARTMAP for ....

G. A. Carpenter and A.-H. Tan. Rule extraction: From neural architecture to symbolic representation. Connection Science, 7(1):3--27, 1995.

....the cluster layer encodes a pair of input and output template patterns. Learned weight vectors, one for each cluster node, thus correspond to a set of rules that link antecedents to consequents. This knowledge structure allows the translation of the network architecture to a set of symbolic rules [1]. In addition, symbolic rules can be inserted into an inductive NLN before learning and refined using the SCM learning algorithm. Integration of symbolbased and pattern cased knowledge will be our next research focus. ....

G. A. Carpenter and A. H. Tan. Rule extraction: From neural architecture to symbolic representation. Connection Science, 7:in press, 1995.

No context found.

G. Carpenter and T. A., \Rule extraction: From neural architecture to symbolic representation," Connection Science, vol. 7, no. 1, pp. 3-27, 1995.

No context found.

G. A. Carpenter and A. H. Tan, "Rule extraction, from neural architecture to symbolic representation," Connection Sci., vol. 7, pp. 3--27, 1977.

No context found.

Carpenter, A and Tan, A H, "Rule extraction: from neural architecture to symbolic representation" Connection Science Vol 7 No 1 (1995) pp 3-27

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