- 1-1 Introduction
- 1-2 Decomposition and Reconstruction
- 1-3 Neurons and Synapses
- 1-4 Neural Networks
- 1-5 Systems Control Mechanisms in the CNS
- 1-6 Reflexes and Voluntary Movements
- 1-7 Integration of reflexes
- 1-8 Motor Actions
- 1-9 Cognitive Functions
- 1-10 Beyond Movements
- 1-11 Scope of This Monograph
- 1-12 Summary
This chapter is from the book
This chapter is from the book
This chapter is from the book
1-4 Neural Networks
Numerous neurons in the CNS assemble to form a structure called a "nucleus." In certain areas of the brain and spinal cord (e.g., the superior colliculus, cerebellar cortex, hippocampal cortex, cerebral neocortex), different types of neurons regularly assemble to form a multilayered network. Donald Hebb (1904–1985) proposed the concept of "neuron assembly," that is, a collation of neurons interconnected by synapses, in which the connectivity is modifiable according to experienced activities (Hebb, 1949). A famous proposal by Hebb is that the connection between two neurons firing synchronously is strengthened. Because of this "Hebbian" type of synaptic plasticity; a neuronal assembly can change its circuitry structure (corresponding to memory) and consequently modify its input-output relationships (corresponding to learning), as dependent on experienced activities.
In an effort to verify Hebb's concept of neuron assembly, Frank Rosenblatt (1928–1971) constructed a model network named a "simple perceptron." It consisted of three layers of neurons connected in one direction, from the sensory cell layer to the association cell layer, to the response cell layer (Figure 6). Connections from the first to the second layer were fixed, whereas those from the second to the third layer were modifiable according to the instruction of an outside "teacher." The teacher increased the weight of connection at all junctions transmitting signals from the second to the third layer when the simple perceptron responded correctly to sensory stimuli. The teacher decreased the weight at all second-to-third layer connections transmitting signals when the response was incorrect. When this training process was repeated, the simple perceptron improved its performance toward a success rate of 100%. This was the first man-made machine capable of learning. Ten years later, a counterpart of the simple perceptron was found in the cerebellum (see Chapters 3 and 9). The simple perceptron exemplified the success of the constructive approach (i.e., to understand by construction) for clarifying the operation of neuronal networks in the CNS.
)
Figure 6 The simple perceptron model of the cerebellum.
This figure is self-explanatory. See the text for further details on the operation of a simple perceptron. Abbreviation: CF, climbing fiber.
Twenty-four years after the construction of the simple perceptron, another form of multilayered neuronal assembly was proposed. It is usually called a "neurocomputer" (Rumelhart et al., 1986), in which errors were estimated by comparing the output of the third layer with a preset goal and were back propagated to the third-layer neurons. The errors acted on the junctions on the third-layer neurons formed with second-layer (hidden layer) neurons, and modified the efficacy of transmission from second-layer to third-layer neurons. The neurocomputer is often applied to model information processing in hippocampal and neocortical networks.