Marin Marinov

The dissertation represents a critical evaluation of the major connectionist theories of human cognitive architecture. The central connectionist thesis that artificial neural networks can serve as finitary models of human cognizers is examined and rejected. Connectionist theories, in contrast to the classical symbol-processing theories of cognitive architecture, cannot explain the productivity and systematicity of mental states. The reason for this is that ANN-based cognitive architectures cannot maintain representational states with compositional structure. Chapter One analyzes the implementational connectionism's solution to the problem of compositionality. It is shown that neither the theory of weak nor of strong compositionality can solve this problem. ;Chapter Two criticises the attempt to establish connectionism as an alternative theory of human cognitive architecture through the introduction of the symbolic/subsymbolic distinction. The reasons for the introduction of this distinction are examined and found to be unconvincing. Several experimental comparisons between the TDIDT class of symbolic learning systems and the class of artificial neural networks using the error backpropagation algorithm are discussed. It is argued that the differences in the performance of these two classes of learning systems are insignificant and are not systematic. Such evidence contradicts the view that ANNs define a new kind of "subsymbolic" computation. ;Supporters of eliminative connectionism have argued for a pattern association and pattern recognition-based explanation of cognitive processes. They deny that explicit rules and symbolic representations play any role in cognition. Their argument is based to a large extent on Rumelhart and McClelland's and MacWhinney and Leinbach's connectionist models of learning of the past tenses of English verbs. Chapter Three presents an analysis of an experimental comparison between these models and the Symbolic Pattern Associator --a learning system based on the classical architecture. It is shown that the SPA outperforms the connectionist models; moreover, the SPA can represent the acquired knowledge in the form of explicit rules. The analysis of this comparison leads to the conclusion that symbol-processing models have a far better chance of explaining complex cognitive phenomena in terms of rules and symbolic representations than eliminative connectionism

Quantal penetration through a (stationary) one-dimensional potential barrier is considered as a time evolution of an initially prepared wave packet. The large-time asymptotics of the process is concerned. Locality of the potential imposes certain analytical properties of the interaction amplitudes in the energy representation. The results are presented in terms of development of the phase-space (Wigner's) quasi-distribution. The phase-space evolution kernel is constructed, and it is shown that in the presence of a positive potential no part of the distribution is transported faster than the free particle. For the case of a small initial momentum uncertainty, the deformation of the coordinate density is considered, including a possible advance of its maximum, which would not mean any noncausal signal transport

This article presents counter evidence against Smolensky's theory that human intuitive/nonconscious congnitive processes can only be accurately explained in terms of subsymbolic computations carried out in artificial neural networks. We presentsymboliclearning models of two well‐studied, complicated cognitive tasks involving nonconscious acquisition of information: learning production rules and artificial finite state grammars. Our results demonstrate that intuitive learning does not imply subsymbolic computation, and that the already well‐established, perceived correlation between “conscious” and “symbolic” on the one hand, and between “nonconscious” and “subsymbolic” on the other, does not exist.

The article criticises the attempt to establish connectionism as an alternative theory of human cognitive architecture through the introduction of thesymbolic/subsymbolic distinction (Smolensky, 1988). The reasons for the introduction of this distinction are discussed and found to be unconvincing. It is shown that thebrittleness problem has been solved for a large class ofsymbolic learning systems, e.g. the class oftop-down induction of decision-trees (TDIDT) learning systems. Also, the process of articulating expert knowledge in rules seems quite practical for many important domains, including common sense knowledge.The article discusses several experimental comparisons betweenTDIDT systems and artificial neural networks using the error backpropagation algorithm (ANNs usingBP). The properties of one of theTDIDT systemsID3 (Quinlan, 1986a) are examined in detail. It is argued that the differences in performance betweenANNs usingBP andTDIDT systems reflect slightly different inductive biases but are not systematic; these differences do not support the view that symbolic and subsymbolic systems are fundamentally incompatible. It is concluded, that thesymbolic/subsymbolic distinction is spurious. It cannot establish connectionism as an alternative cognitive architecture.

Supporters of eliminative connectionism have argued for a pattern association-based explanation of language learning and language processing. They deny that explicit rules and symbolic representations play any role in language processing and cognition in general. Their argument is based to a large extent on two artificial neural network (ANN) models that are claimed to be able to learn the past tenses of English verbs (Rumelhart & McClelland, 1986, Parallel distributed processing, Vol. 2, Cambridge, MA: MIT Press; MacWhinney & Leinbach, 1991, Cognition, 40, 121-157). In this article we critically review Rumelhart and McClelland's as well as MacWhinney and Leinbach's ANN models and conclude that they do not succeed in the assigned task of learning the past tenses of English verbs. In order to answer their challenge to the symbolic processing approach, we present our symbolic pattern associator (SPA)-a general-purpose pattern associator that can learn to associate arbitrary discrete patterns. We carried out several experiments with the SPA using the same set of verbs that was used in MacWhinney and Leinbach's simulation with more realistic training and testing procedures. The SPA outperformed the connectionist models by a wide margin in the accuracy of learning, and successful inductive generalizations to unseen verbs. Our SPA has very natural and psychologically realistic explanations to many psychological effects such as U-shaped learning curve, and is much closer to human subjects in predicting past tense of the pseudo-verbs. In contrast to ANNs, whose internal representations are entirely opaque, the SPA can represent the acquired knowledge in the form of production rules that allow for further higher-level processing and integration, resulting in linguistically realistic associative templates for irregular verbs and production rules for regular verbs. In the light of these findings, we conclude that eliminative connectionists' vision of cognition as simple pattern association and pattern recognition without symbolic representation is inadequate. Pattern association as such does not imply rule-less or cue-based models of language acquisition or of human learning in general.