06.09.2021 Views

Mind, Body, World- Foundations of Cognitive Science, 2013a

Mind, Body, World- Foundations of Cognitive Science, 2013a

Mind, Body, World- Foundations of Cognitive Science, 2013a

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Perceptrons, too, can match probabilities (Dawson et al., 2009). Dawson et al.<br />

used four different cues, or discriminative stimuli (DSs), but did not “reward” them<br />

100 percent <strong>of</strong> the time. Instead, they rewarded one DS 20 percent <strong>of</strong> the time,<br />

another 40 percent, a third 60 percent, and a fourth 80 percent. After 300 epochs,<br />

where each epoch involved presenting each cue alone 10 different times in random<br />

order, these contingencies were inverted (i.e., subtracted from 100). The dependent<br />

measure was perceptron activity when a cue was presented; the activation function<br />

employed was the logistic. Some results <strong>of</strong> this experiment are presented in Figure<br />

4-6. It shows that after a small number <strong>of</strong> epochs, the output unit activity becomes<br />

equal to the probability that a presented cue was rewarded. It also shows that perceptron<br />

responses quickly readjust when contingencies are suddenly modified, as<br />

shown by the change in Figure 4-6 around epoch 300. In short, perceptrons are<br />

capable <strong>of</strong> probability matching.<br />

Mean Network Activation<br />

Epoch <strong>of</strong> Training<br />

Figure 4-6. Probability matching by perceptrons. Each line shows the perceptron<br />

activation when a different cue (or discriminative stimulus, DS) is presented.<br />

Activity levels quickly become equal to the probability that each cue was<br />

reinforced (Dawson et al., 2009).<br />

That perceptrons match probabilities relates them to contingency theory. Formal<br />

statements <strong>of</strong> this theory formalize contingency as a contrast between conditional<br />

probabilities (Allan, 1980; Cheng, 1997; Cheng & Holyoak, 1995; Cheng & Novick,<br />

1990, 1992; Rescorla, 1967, 1968).<br />

For instance, consider the simple situation in which a cue can either be presented,<br />

C, or not, ~C. Associated with either <strong>of</strong> these states is an outcome (e.g., a<br />

reward) that can either occur, O, or not, ~O. In this simple situation, involving a<br />

single cue and a single outcome, the contingency between the cue and the outcome<br />

is formally defined as the difference in conditional probabilities, P, where P =<br />

P(O|C) – P(O|~C) (Allan, 1980). More sophisticated models, such as the probabilistic<br />

contrast model (e.g., Cheng & Novick, 1990) or the power PC theory (Cheng, 1997),<br />

154 Chapter 4

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!