Online research in philosophy

The importance of the Stability Problem in neurocomputing is discussed, as well as the need for the study of infinite networks. Stability must be the key ingredient in the solution of a problem by a neural network without external intervention. Infinite discrete networks seem to be the proper objects of study for a theory of neural computability which aims at characterizing problems solvable, in principle, by a neural network. Precise definitions of such problems and their solutions are given. Some consequences are explored, in particular, the neural unsolvability of the Stability Problem for neural networks.

Neural organization achieves its stated goal to “show how theory and experiment can supplement each other in an integrated, evolving account of structure, function, and dynamics” (p. ix), showing in a variety of contexts – from olfactory processing to spatial navigation, motor learning and more – how function may be realized in the neural tissue, with explanatory and predictive neural network models providing a cornerstone in this approach.

In "Brainshy: Non-neural theories of conscious experience," (this volume) Patricia Churchland considers three "non-neural" approaches to the puzzle of consciousness: 1) Chalmers' fundamental information, 2) Searle's "intrinsic" property of brain, and 3) Penrose-Hameroff quantum phenomena in microtubules. In rejecting these ideas, Churchland flies the flag of "neuralism." She claims that conscious experience will be totally and completely explained by the dynamical complexity of properties at the level of neurons and neural networks. As far as consciousness goes, neural network firing patterns triggered by axon-to-dendrite synaptic chemical transmissions are the fundamental correlates of consciousness. There is no need to look elsewhere.

Synchronization of neural activity in oscillatory neural networks is a general principle of information processing in the brain at both preattentional and attentional levels. This is confirmed by a model of attention based on an oscillatory neural network with a central element and models of feature binding and working memory based on multi-frequency oscillations.

Artificial neural networks have weaknesses as models of cognition. A conventional neural network has limitations of computational power. The localist representation is at least equal to its competition. We contend that locally connected neural networks are perfectly capable of storing and retrieving the individual features, but the process of reconstruction must be otherwise explained. We support the localist position but propose a “hybrid” model that can begin to explain cognition in anatomically plausible terms.

Body representations traverse the whole of the brain. They provide vital sources of information for every facet of an animal’s behavior, and such direct neural connectivity of visceral input throughout the nervous system demonstrates just how strongly cognitive systems are linked to bodily representations. At each level of the neural axis there are visceral appraisal systems that are integral in the organization of action. Cognition is not one side of a divide and viscera the other, with action merely a reflexive outcome. There is no divide between cognition and bodily functions once the brain is involved. Cognitive mechanisms that permeate neural function are a cardinal piece of biological function and adaptation.

The present commentary addresses the Quartz & Sejnowski (Q&S) target article from the point of view of the dynamical learning algorithm for neural networks. These techniques implicitly adopt Q&S's neural constructivist paradigm. Their approach hence receives support from the biological and psychological evidence. Limitations of constructive learning for neural networks are discussed with an emphasis on grammar learning.

Both cognitive science and phenomenology accept the primacy of the organism-environment system and recognize that cognition should be understood in terms of an embodied agent situated in its environment. How embodiment is seen to shape our world, however, is fundamentally different in these two disciplines. Embodiment, as understood in cognitive science, reduces to a discussion of the consequences of having a body like ours interacting with our environment and the relationship is one of contingent causality. Embodiment, as understood phenomenologically, represents the condition of intelligibility of certain terms in our experience and, as such, refers to one aspect of that background which presupposes our understanding of the world. The goals and approach to modeling an embodied agent in its environment are also fundamentally different dependent on which relationship is addressed. These differences are highlighted and are used to support our phenomenologically based approach to organism-environment interaction and its relationship to brain function.

Borrett, Kelly and Kwan [(2000) Phenomenology, dynamical neural networks and brain function, Philosophical Psychology, 13, 000-000] claim that unbiased, self-evident, direct description is possible, and may supply the data that brain theories account for. Merleau-Ponty's [(1962) Phenomenology of perception, London: Routledge] description of Schneider's apraxia is offered as a case in point. According to the authors, Schneider's apraxia justifies brain components of predicative and pre-predicative experience. The description derives from a bias, however, that parallels modularity's morphological reduction. The presence of biasing presuppositions contradicts the goal of direct description. Moreover, the authors' brain account is not necessary to explain Schneider's apraxia, and morphological reduction is not sufficient to explain emergent phenomena of motor control.

I address whether neural networks perform computations in the sense of computability theory and computer science. I explicate and defend
the following theses. (1) Many neural networks compute—they perform computations. (2) Some neural networks compute in a classical way.
Ordinary digital computers, which are very large networks of logic gates, belong in this class of neural networks. (3) Other neural networks
compute in a non-classical way. (4) Yet other neural networks do not perform computations. Brains may well fall into this last class.