Toward a Noncomputatiod Cognitive Neuroscience Gordon G. Globus University of California Irvine Abstract Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/4/4/299/1754995/jocn.1992.4.4.299.pdf by guest on 18 May 2021 The near universally accepted theory that the brain pro- information processing is developed in which the brain permits cesses information persists in current neural network theory and supports “participation” of self and other as constraints on where there is “subsymbolic” computation (Smolensky, 1988) the dynamically evolving, self-organizing whole. The noncom- on distributed representations. This theory of brain information putational process of “differing and deferring” in nonlinear processing may suffice for simplifying models simulated in dynamic neural systems is contrasted with Black’s (1991) ac- silicon but not for living neural nets where there is ongoing count of molecular information processing. State hyperspace chemical tuning of the input/output transfer function at the for the noncomputational process of nonlinear dynamical sys- nodes, connection weights, network parameters, and connec- tems, unlike classical systems, has a fractal dimension. The tivity. Here the brain continually changes itself as it intersects noncomputational model is supported by suggestive evidence with information from the outside. An alternative theory to for fractal properties of the brain. INTRODUCTION neural sciences is so widespread and automatic that it begins to appear less like an innovative leap than The theory that the brain performs computations is so like a bandwagon phenomenon, of the sort often ob- widespread in cognitive neuroscience, and so strongly served in the sociology and history of science. held, that its status is “paradigmatic” in Kuhn’s (1976) (Daugman, 1990, p. 15) strong sense. It is difficult to even think of brain com- putation as a “theory” until alternatives are seen, so Emboldened by Daugman’s caveat, an alternative theory deeply rooted is the information processing idea in cur- that brain functioning is not computational is proposed rent thought. Thus this journal’s stated purpose is to here. This alternative theory is a “radical” form of con- publish papers that “bridge the gap between descriptions nectionism (neural net theory). of information processing and specifications of brain The alternative is labeled “radical” because current activity.” Indeed “computational neuroscience” is consid- connectionism fully accepts the computational view. The ered to be a subdiscipline (Schwartz, 1990), whose “ul- neural nets of the brain are performing computations, it timate aim” is to “explain how electrical and chemical is thought, though not like desktop computers do. There signals are used in the brain to represent and process is no programmed logical processing of information in information” (Sejnowski, Koch, & Churchland, 1988, real neural nets; the processing is differently performed. p. 1299). But nonetheless there is a computation going on, a “sub- Daugman (1990) reminds us, however, that computa- symbolic” computation, as Smolensky (1988) argues, not tion is a metaphor of brain function. a symbolic one, as Fodor and Pylyshn (1988) think. This computation is naturally expressed by the mathematics Surprisingly, given the pervasive popularity of this of linear algebra (Jordan, 1986), in which the computa- metaphor, there remains today no well-established tion is thought of as the mapping of an input vector to evidence of symbolic manipulation or formal logical an output vector. rules at the neurobiologicdl level in animal physiol- The meaning of “computation” requires some com- ogy. while the computational metaphor often ment, since it is so central to what follows. In their seems to have the status of an established fact, it discussion, “What is computational neuroscience?,” should be regarded as an hypothetical, and historical, Churchland, Koch, and Sejnowski (1990) consider a conjecture about the brain. Today’s embrace of physical system “in a most general sense” to compute a the computational metaphor in the cognitive and hnctionflx) when there is 0 1992 Massachusetts Institute of Technology Journal of Cognitive Neuroscience Volume 4, Number 4 Downloaded from http://www.mitpressjournals.org/doi/pdf/10.1162/jocn.1992.4.4.299 by guest on 23 September 2021 (1) a mapping between the system’s physical inputs logical vacuum” and computational models that are “par- and x, (2) a mapping between the system’s physical ticularly impoverished compared to biologically based outputs andy, such that (3)f(x) = y. (p. 48) brain function” (p. xii). Shepherd (1990) observes, This is a very general account of cornputxion in terms One must conclude that there is at present a consid- of competence, and a “weak” sense of computation. But erable mismatch between neural networks and the they also give “a suitably broad notion of computation” real nervous system. A case could very well be made in terms of perjGormance. that the networks could hardly be less neural; that most of the properties believed to be neural are in It is useful , . to class certain functions in the brain fact not so or are only remotely so, and that many as computational because nervous systems represent properties are introduced that are explicitly not and they respond on the basis of the representations. neural. (p. 92) Nervous systems represent the external world, the body they inhabit, and, in some instances, parts of Despite these strong doubts on the neurobiological side, Downloaded from http://mitprc.silverchair.com/jocn/article-pdf/4/4/299/1754995/jocn.1992.4.4.299.pdf by guest on 18 May 2021 the nervous system itself. (p. 48) neural net modelers remain confident that The performance of the computation makes uses of rep- if the computational model is firmly based on the resentations. Furthermore, the explanation of state tran- available experimental data, it can evolve along with sitions in the performance of the computation “describes the experimental program and help to guide future the states in terms of the information transformed, rep- research directions. (Sejnowski, Koch, & Churchland, resented and stored” (p. 48). So “computation” in the 1988, p. 1305) strong sense of performance is tantamount to informa- tion processing, whereas in the weak sense the term We shall see below that simplifying models serve to sustain the theory that the brain computes whereas bi- vacuously skirts the issue of how the computation is done ologically realistic models open toward the possibility of and thus is divorced from neuroscience. The term “com- a noncomputational neuroscience. putation” accordingly will be used in the strong perfor- Simplified nets consist of formal neurons or richly mance sense in what follows. interconnected “nodes.”Signals fan in on the nodes and This presentation of a radical connectionism in which signals fan out from the nodes; the nodes open both the brain does not actually perform computations begins ways to the whole. There are input/output transfer func- with a critique of simplified neural net models. It is tions at the nodes. Connections between nodes (“con- shown that simplified silicon nets can be thought of as nectivity”) are excitatory or inhibitory and there are computing but biologically realistic nets are noncom- variable weights on the connections. There are also net- putational. Rather than structure sensitive rule governed work parameters such as signaunoise ratio, excitation/ operations on symbolic representations, there is an evo- inhibition ratio, and decay time that affect signal pro- lution of self-organizing nonlinear dynamic systems in a cessing. The net is layered into an input layer on which process of “differing and deferring.” Then a comparison an activation pattern is impressed from the outside, a is made between a biologically realistic molecular infor- hidden layer where distributed processing occurs, and mation processing model (Black, 1991) and the noncom- an output layer that transduces signals to the outside. putational model. Finally, studies of hypothesized fractal During the operation of simplified nets, the connec- properties of the brain are considered that provide some tivity, the connection weights, transfer functions at the preliminary empirical support for the noncomputational nodes, parameters, and input are kept constant. The net model. evolves as a dynamic system and settles into a state that optimizes satisfaction of the various constraints; the set- SIMPLIFIED NEURAL NETS tlement is self-consistent. This process is not pro- grammed but is self-organizing. The settling time Simplification of neural net models is thought to be demarcates an event or “interval.” When nets learn, the unproblematic and consistent with strategies successfully connection weights are changed before the next trial; applied elsewhere in science. Simplifying models “ab- the weight changes constitute a memory trace. Connec- stract from the complexity of individual neurons and the tivity, transfer functions, and network parameters are typ- patterns of connectivity in exchange for analytic tract- ically held constant from trial to trial; their settings can ability” (Sejnowski, Koch, & Churchland, 1988, p. 1301). be thought of as static traces. Connection weights mod- Crick (1989), in contrast, is dubious about the simplifying ified by learning provide dynamic traces in the simplified approach. “Most of these neural ‘models’ are not . model. really models at all, because they do not correspond