Date of Award
Doctor of Philosophy in Psychology
Charles E. Collyer
This experiment was designed to examine the possible physiological mechanisms underlying the formation of critical periods in development. It was hypothesized that experientially-modulated neuronal death, coupled with an overconnected neural network, and acting early in the developmental sequence, could produce a learning effect similar to that described as a "critical period effect."
A computer-based associative memory model was used to simulate a simple neural network representing the associative cortex. Three different independent variables were manipulated in the experiment. The first was the type of connectivity matrix used when the model was initialized prior to training.
Four different developmental paradigms were modeled. One matrix simulated a neural structure with specific, genetically-determined characteristics. A second matrix simulated a "tabula rasa" or blank slate paradigm. A third matrix simulated an initially overconnected network. The last was a randomly generated connectivity matrix that reflected no particular paradigm, used as a control.
A second independent variable was the experiential sequence in which the model was trained. Each developmental model was trained on four different stimulus sets simulating different experiential sequences.
The third independent variable was the presence or absence of an experientially-modulated neuronal death component. All developmental models were trained across all experiential sequences with and without the neural death component acting on the model.
The model's behavior was assessed using 2 dependent variables - the ability of the model to learn later in development, and the structural fit of the model to the prototype stimulus set. In addition, the model's output was examined for evidence of a critical period effect.
The results of the experiment indicate that a critical period effect can be induced simply by the use of an
experientially-modulated neural death, operating on a plastic neural structure. A genetically-determined structure is not necessary for the manifestation of critical period phenomena. All models performed similarly. It would appear that a hardwired, genetically-determined neural structure may not be the optimum paradigm to support later learning in a noisy environment, that is, one where the environmental stimuli differ in some degree from those that the model has been structured to deal with.
Surprisingly, both the ability to learn and the structural fit of the overconnected and experiential models were significantly better than that of the genetic model in noisy environments. Only the randomly-connected model performed worse than the genetic model. It appears that genetically specifying a neural structure that must respond to a noisy environment produces problems similar to those of overfitting an analytical model.
The ability of all models to learn a specific stimulus was expected to increase as exposure to the stimulus increased. This behavior was, in fact, observed for all models.
The findings of this experiment indicate the formation of critical periods may be the natural consequence of rapidly removing unused neural pathways early in development, while stabilizing the pathways that are used. If the organism is not exposed to certain types of stimuli, the neural pathways associated with processing those stimuli are lost. If the organism is exposed to those stimuli early in development, the circuits are stabilized, and are available for learning later in development.
Skrmetti, Thomas J., "Experientially-Modulated Neuronal Death in Overconnected Neural Networks" (1991). Open Access Dissertations. Paper 827.