This is a precursor of the PDP model which, after some modifications, became the Computational DO Model.
This start of the model is the dimensional overlap taxonomy. The events generating Reaction Times are assumed to take place in a network consisting of multiple parallel processing layers (Input – Intermediate – and Output). Earlier layers may send continuous, partial activation to later layers, continuously. Thus, unlike the original DO model and its later computational form, this PDP version of the DO model is not a discrete stage model. While it permits information transmission within the same layer, it does not permit feedback from a later to an earlier layer.
A stimulus or response dimension (e.g. color) is represented by a module which is made up of neuron-like nodes. Nodes represent a stimulus or response feature. The number of nodes within a module is determined by the number of such features (e.g. particular colors, such as red, or blue…). Feature nodes within the same module are negatively connected, mutually inhibitory, and so weighted as parameters.
Modules are arranged in three layers. The input layer represents physical, carrier-specific stimuli; e.g. colored ink, or color words. The intermediate layer represents abstract concepts; e.g. color, location. The output layer represents responses; e.g. color names, key presses.
Modules and nodes in the input layer receive their input from the external environment via “task” lines. “Task lines” represent the attention allocated to different stimulus dimensions, and is thus one of the parameters in the model. The nodes in the input layer generate and send activation to the corresponding nodes in the intermediate layer via “carrier lines”. “Carrier lines” represent the strength between carrier-specific stimuli and their concepts, and is another parameter in the model. The nodes in the intermediate layer produce and send activation to corresponding nodes in the output layer via “control” lines. “Control lines’ represent the SR mapping, i.e. the controlled processes postulated in the DO model. All connections are between modules in different layers.
Because stimuli and responses may be multidimensional, multiple modules may exist in each layer. For SS overlap, (e.g. Ensembles 4 and 8,) the input layer may have two modules: one for the relevant stimulus, the other for the irrelevant stimulus. The model assumes that these two modules converge on a common module in the intermediate level.
For SR overlap (e.g. Ensemble 2) the corresponding nodes in the intermediate and output layers are linked via “automatic line”, (which is an implementation of “automatic activation” in the original DO model).
The connection patterns for congruent and incongruent mappings are quite similar: corresponding nodes in the intermediate and output layers are connected via automatic lines; however, for incongruent mappings, instead of connecting the control line to the same nodes, it connects to different nodes, thus producing response competition – just as in the original DO model.
In Ensemble 3 the stimulus is two-dimensional. Both the input layer and the intermediate layers each have two modules (relevant and irrelevant). Since the irrelevant dimension overlaps with the response, the nodes in the irrelevant dimension are connected to the corresponding nodes in the output layer via automatic lines. The nodes in the relevant dimension are connected via control lines thus, once again, generating response competition.
With these “architectural principles” now in place, a PDP network for ensembles 1 to 8 may be constructed. Such networks will necessarily differ from each other architecturally when they differ in overlap, or in mapping.
The presentation of a stimulus feature is assumed to activate the corresponding node in the input layer with a value of 1, which is then clamped down, otherwise it remains at 0. These values are fed continuously to the appropriate feature nodes in the output layer. Once the activation of any node in the output layer reaches the response threshold, an overt response is executed.
With some additional processing assumptions (see Zhang, et. al, 1999), and the setting of a number of parameters (including: threshold; the weight for carrier line, control line, and automatic line; mutual inhibition) successful simulations were run with this model on the data of several of our experiments.
The Computational Dimensional Overlap Model incorporates many of the features of this model, but also changes some features of this model to bring the computational model more in line with the original process model. For example, this model is a continuous processing model that does not process stimulus and response information in discrete stages, like the original process model describes. The final implementation of the computational model keeps the network-activation principles of this model while also implementing the property of stage-like processing.