Question
Answered step-by-step
PrivateWater38693
Explain negative patterning and the rabbit experiment Negative…
Explain negative patterning and the rabbit experiment
Negative Patterning:
When the Whole Means Something Different Than the Parts
In another complication of generalization from compounds composed of stimuli with different attributes, consider what happens when Sabrina starts playing with other children. She might learn from experience on the playground that freckle-faced children are usually bratty and that red-haired children are also bratty. Given these two associations, you might expect her to be especially wary of a girl she meets one day who is both red haired and freckle faced. Based on her past experiences, she expects the girl to be really horrid because she has both of the attributes Sabrina associates with brattiness. All other things being equal, we, like Sabrina, tend to assume that combinations of cues will have consequences that combine, and possibly summate, what is known about the individual cues.
But is this assumption always correct? What if a certain combination of cues implies something totally different from what the individual cues mean? For example, you know what the letter “c” sounds like, and you know what the letter “h” sounds like, but when you see these two letters together as “ch,” you know that they represent quite a different sound. Thus, readers of English learn that certain combinations of letters can sound very different from their component letters. Consider another example, involving driving a car. While waiting at an intersection, you notice that the antique Citroën car in front of you has its left rear red taillight flashing, signaling the car is about to turn left (Figure 6.7A). If its right rear red taillight were flashing, you would assume the driver intends to turn right (Figure 6.7B).
FIGURE 6.7 The challenge of interpreting combined cues (A) Left blinking light means left turn, (B) right blinking light means right turn, (C) both lights blinking means that the driver has turned on the hazard lights.
Photo A shows a red star on the left taillight. Photo B shows a red star on the right taillight. Photo C shows red stars on the right and left taillights.
But what if both taillights are flashing, as in Figure 6.7C? Although either the left or the right turn signal flashing indicates an imminent turn, both taillights flashing together certainly does not indicate a combination of two turns. Instead, twin flashing taillights signal a hazard: the car is proceeding slowly or is disabled. Although our tendency may be to assume that what is true of component features presented individually is also true of their combination, clearly it is possible to override this tendency to generalize from components to compounds, as automobile drivers who learn that the combination of both left and right lights flashing means something quite different than either of them flashing alone.
This kind of situation, where cue combinations have radically different meanings than their components, has been extensively studied with both animal and human learning tasks. Suppose, for example, that a rabbit in a classical conditioning study is trained to expect that either a tone or a light, presented alone, predicts an airpuff US to the eye but that there will be no airpuff US if the tone and light appear together. To respond appropriately, the animal must learn to respond with a blink to the individual tone cue or light cue but to withhold responding to the compound tone-and-light cue. This task, schematized as
Tone ? airpuff US
Light ? airpuff US
Tone + light ? no US
is known as negative patterning because the response to the individual cues is positive, whereas the response to the compound (i.e., the “pattern”) is negative (no response).
negative patterning
A behavioral paradigm in which the appropriate response to individual cues is positive, whereas the appropriate response to their combination (pattern) is negative (no response).
Because both the tone and the light cues are part of the tone-and-light compound, there is a natural tendency for the animal to generalize from the component cues to the compound cue and vice versa. However, this natural tendency to generalize from component features to compounds does not work in this case because the components and the compound are associated with very different outcomes.
With training, the negative-patterning task can be mastered by many animals, including humans. Figure 6.8 shows an example of negative patterning in rabbit eyeblink conditioning (Kehoe, 1988). After only a few blocks of training, the animals learn to give strong responses both to the tone alone and to the light alone. In addition, during this early phase of training, the rabbits make the mistake of overgeneralizing from the components to the compound, giving incorrect strong responses to the compound tone-and-light stimulus (which is not followed by an airpuff). Only with extensive further training do the rabbits begin to suppress responding to the tone-and-light compound.
FIGURE 6.8 Negative patterning in rabbit eyeblink conditioning Negative patterning requires learning to respond to two cues (a tone and light) when each is presented separately but to withhold the response when the two cues are presented together.
Data from Kehoe, 1988.
The horizontal axis is labeled blocks of trials, ranging from 0 to 10, in increments of 1. The vertical axis is labeled percent responding, ranging from 0 to 100, in increments of 25. A line labeled tone starts at (1, 5) and rises through (2, 10), (3, 35), (4, 40), (7, 80), (10, 70), (11, 75), and (14, 75), before ending at (18, 85). A line labeled light starts at (1, 0) and rises through (3, 35), (4, 40), (6, 75), (8, 80), (9, 75), (13, 75), and (17, 80), before ending at (18, 70). A line labeled tone plus light starts at (1, 10) and rises through (2, 20), (3, 50), (6, 80), (7, 75), (8, 80), (12, 75), (13, 60), (14, 55), (15, 40), (16, 39), and (17, 35), before ending at (18, 40). All values are estimated.
Can associative-network learning models such as those illustrated in Figure 6.2 (with discrete-component representation) or Figure 6.4 (with distributed representations to indicate stimulus similarity) provide an explanation of how rabbits and other animals learn negative patterning? Figure 6.9 illustrates why single-layer network models, like those in Figure 6.2, that use discrete-component representations cannot learn the negative-patterning problem. To produce correct responding to the tone alone, the modifiable associative weight from the input unit encoding tone must be strengthened to 1.0 (Figure 6.9A). To produce correct responding to the light alone, the weight from the input unit encoding light must also be strengthened to 1.0 (Figure 6.9B). But this means that if the tone and light are presented together, activation will both flow through those modifiable weighted connections and produce strong responding to the compoundstronger responding, in fact, than to either component alone (Figure 6.9C).
FIGURE 6.9 Failure of a single-layer network with discrete-component representations to learn negative patterning (A) For the tone cue to correctly generate a strong response, the connection from that input to the output must be strongly weighted. (B) For the light cue to correctly generate a strong response, the connection from that input to the output must also be strongly weighted. (C) Consequently, when both tone and light cues are present, the network will incorrectly give a strong response.
All three diagrams have two input nodes each. Each input node is represented by a circle with a number inside. The input nodes are placed side by side in a row. The input nodes from left to right are tone and light. An output node (represented by a circle with a number inside) is above the space between the two input nodes. Associative weights are arrows which point from the input nodes to the output nodes. In diagram A, tone alone, the input node for tone is numbered 1 and is a red circle. The input node for light is numbered 0. The associative weight arrow for input node tone is a red thick arrow labeled, 1.0. The associative weight arrow for input node light is a gray thick arrow labeled, 1.0. In diagram B, light alone, the input node for light is numbered 1 and is a red circle. The input node for tone is numbered 0. The associative weight arrow for input node light is a red thick arrow labeled, 1.0. The associative weight arrow for input node tone is a gray thick arrow labeled, 1.0. In diagrams A and B, the output node circle is numbered 1 and is red . An upward arrow from output node points to the text, response. In diagram C, tone plus light, both the input nodes are numbered 1 and are red circles. The associative weight arrows for the input nodes are red thick arrows labeled, 1.0. The output node circle is numbered 2 and is red . An upward arrow from output node points to the text, response.
All the weights could be decreased, of course, to reduce the level of response to the compound in Figure 6.9C, but this would also, incorrectly, reduce responding to the individual components. In fact, there is no way to assign associative weights in the network of Figure 6.9 that would make the network respond correctly to all three different types of training trials.
Negative patterning is just one example of a larger class of learning phenomena that involve configurations of stimuli and that cannot be explained using single-layer networks. To master such tasks, an animal must be sensitive to the unique configurations (or combinations) of the stimulus cues, above and beyond what it knows about the individual stimulus components. Solving such problems requires more complex networks than the single-layer network of Figure 6.9either a single-layer network with more complex nodes that are sensitive to configurations of combinations of cues or a network with multiple layers of weights and nodes that can learn complex cue combinations.
