4.17: Connectionist Reorientation
- Page ID
In the reorientation task, an agent learns that a particular place—usually a corner of a rectangular arena—is a goal location. The agent is then removed from the arena, disoriented, and returned to an arena. Its task is to use the available cues to relocate the goal. Theories of reorientation assume that there are two types of cues available for reorienting: local feature cues and relational geometric cues. Studies indicate that both types of cues are used for reorienting, even in cases where geometric cues are irrelevant (Cheng & Newcombe, 2005). As a result, some theories have proposed that a geometric module guides reorienting behaviour (Cheng, 1986; Gallistel, 1990).
The existence of a geometric module has been proposed because different kinds of results indicate that the processing of geometric cues is mandatory. First, in some cases agents continue to make rotational errors (i.e., the agent does not go to the goal location, but goes instead to an incorrect location that is geometrically identical to the goal location) even when a feature disambiguates the correct corner (Cheng, 1986; Hermer & Spelke, 1994). Second, when features are removed following training, agents typically revert to choosing both of the geometrically correct locations (Kelly et al., 1998; Sovrano et al., 2003). Third, when features are moved, agents generate behaviours that indicate that both types of cues were processed (Brown, Spetch, & Hurd, 2007; Kelly, Spetch, & Heth, 1998).
Recently, some researchers have begun to question the existence of geometric modules. One reason for this is that the most compelling evidence for claims of modularity comes from neuroscience (Dawson, 1998; Fodor, 1983), but such evidence about the modularity of geometry in the reorientation task is admittedly sparse (Cheng & Newcombe, 2005). This has led some researchers to propose alternative notions of modularity when explaining reorientation task regularities (Cheng, 2005, 2008; Cheng & Newcombe, 2005).
Still other researchers have explored how to abandon the notion of the geometric module altogether. They have proceeded by creating models that produce the main findings from the reorientation task, but they do so without using a geometric module. A modern perceptron that uses the logistic activation function has been shown to provide just such a model (Dawson et al., 2010).
The perceptrons used by Dawson et al. (2010) used a single output unit that, when the perceptron was “placed” in the original arena, was trained to turn on to the goal location and turn off to all of the other locations. A set of input units was used to represent the various cues—featural and geometric—available at each location. Both feature cues and geometric cues were treated in an identical fashion by the network; no geometric module was built into it.
After training, the perceptron was “placed” into a new arena; this approach was used to simulate the standard variations of the reorientation task in which geometric cues and feature cues could be placed in conflict. In the new arena, the perceptron was “shown” all of the possible goal locations by activating its input units with the features available at each location. The resulting output unit activity was interpreted as representing the likelihood that there was a reward at any of the locations in the new arena.
The results of the Dawson et al. (2010) simulations replicated the standard reorientation task findings that have been used to argue for the existence of a geometric module. However, this was accomplished without using such a module. These simulations also revealed new phenomena that have typically not been explored in the reorientation task that relate to the difference between excitatory cues, which indicate the presence of a reward, and inhibitory cues, which indicate the absence of a reward. In short, perceptrons have been used to create an associative, nonmodular theory of reorientation.