Stimulus Control Part 1

Stimulus Control: Roadmap of Today’s Lecture

This lecture is about how living things learn to react to specific signals in their environment. We'll look at three main ideas:

Discrimination: Learning to react to one signal (S+) but not another similar one (S–). Think of a dog learning to sit only when you say "sit," not when you say "kit."
Generalization: Reacting the same way to signals that are similar but not exactly identical. Like a baby calling all four-legged animals "doggy" at first.
Categorization: Grouping different signals into one mental category. For example, knowing that different breeds of dogs are all still "dogs."

We'll also ask:

How do animals make these connections?
Do they just learn about each signal on its own (absolute), or do they compare signals and learn about their relationships (relational)?

Real-World Showcases of Discrimination, Generalization & Categorization

Tracking dogs: Imagine a dog trying to find a missing child. It's given something with the child's smell. The dog has to:
- Ignore other smells (discrimination).
- Still follow the scent even if it's weaker or mixed with other things (generalization).
- Treat every hint of that child's smell as part of the same task: finding that specific child (categorization).
Skinner’s “Pigeon-guided missile”: During WWII, pigeons were trained to guide missiles by pecking at images of ships on a screen. They had to:
- Peck even if the ship looked different (bigger, smaller, from a different angle) – this is generalization.
- Ignore the background and only focus on the ship – this is discrimination.
- This shows they recognized the category "ship."

Historical Theoretical Debate: Absolute vs. Relational Associations

For a long time, scientists debated how animals learn:

Absolute (mechanistic / behaviorist): This idea says an animal learns about each signal separately. If a signal leads to food, its "strength" increases. If not, it decreases. It's like learning simple a $+$ b $=$ c.
Relational (cognitive / Gestalt): This idea suggests animals understand the relationship between signals. For example, they don't just pick the "bright light"; they pick the brightest out of several lights. This requires comparing things.

Kohler (1918) – Classic Chicken Brightness Transposition Study

How it worked: Chickens were trained with two gray cards.
- S+ (medium-gray) meant food.
- S– (darker-gray) meant no food.
The test: After training, the chickens were shown the original S+ (medium-gray) and a new, even lighter gray card (S'). There was no food given in this test.
Results: The chickens consistently pecked the new, lighter gray card (S'), even though they had never been trained on it specifically. This seemed to show they weren't just picking the medium-gray card; they were picking the lighter of the two cards – a relational rule!

Behaviorist Counter-Attack – Kenneth Spence’s Gradient Theory (1937–40’s)

Spence, a behaviorist, tried to explain Kohler's results without saying the chickens were thinking relationally. His theory said:

Assumptions: When an animal learns about a signal (like a color or sound), it develops a "strength" related to that signal. If the signal leads to food, it generates "excitement" ( $+V$ ). If it leads to no food, it generates "inhibition" ( $-V$ ).
- This strength isn't just for the specific signal; it spreads (or generalizes) to similar signals, like a wave. The closer a new signal is to the original one, the more it's affected.
For Kohler’s experiment: Spence proposed:
- The medium-gray S+ created a "wave" of excitement around itself.
- The darker-gray S– created a "wave" of inhibition around itself.
- When you add these waves together (net strength $V<em>{ ext{net}} = V</em>{ ext{exc}} + V_{ ext{inh}}$ ), the highest point of excitement ends up being beyond the S+ on the lighter side. This is because the S–'s inhibition "pushes" the peak of excitement for S+ a little bit away from it.
Conclusion: So, without any complex thinking, the animal just pecks at the signal that has the highest combined "strength." Spence argued this explained Kohler's findings.

Empirical Test – Hanson (1959) Pigeon Color Discrimination

Hanson tested Spence's idea with pigeons and colors:

Training: S+ was $550\,\ ext{nm }$ (green light); S– was $590\,\ ext{nm }$ (yellow-green light).
Test: Pigeons were shown many different wavelengths (colors).
Findings: The pigeons pecked the most at about $540\,\ ext{nm }$ – a color shifted away from S– and closer to the green side. This peak shift strongly supported Spence's theory that excitation and inhibition gradients sum up.

Two Empirical Problems for Pure Gradient Account

Despite the success of peak shift, there were issues for Spence's simple gradient idea:

No Reversal of Transposition: If chickens in Kohler's setup were shown the S' (lighter gray) card versus an even lighter S'' card, they still chose the lightest one (S''). This shows the relational rule ("choose the lighter") was still at play, not just some fixed peak strength.
Gonzalez (1967) – Primate “Intermediate Size” Task: Monkeys were trained to pick the medium-sized tin out of three (small, medium, large) to get food. When shown a new, larger set of tins, they picked the new middle-sized tin, not the exact size they were trained on. Again, this points to relational processing (picking the "middle one") rather than just remembering one specific size.

Feature-Positive Discrimination (Wagner 1969)

This experiment looked at how rats learned about a tone and a light:

Setup: Rats learned to press a lever for food.
- Tone + Light: Food (S+)
- Tone alone: No food (S–)
Key point: The tone was present in all trials where the rat could get food (50%), but the light only appeared when food was given.

Competing Predictions: Spence vs. Rescorla-Wagner (R-W)

Spence's idea: He would predict that the tone should get some "strength" because it's sometimes paired with food, maybe half the "strength" of the light.
Rescorla-Wagner (R-W) idea: This theory says that when multiple signals are present, they share the "credit" for the outcome (food). If the light is a perfect predictor of food, it will take all the "credit." This would mean the tone's strength would drop to almost zero.

Data

The rats ultimately pressed the lever only when the Light was present. When the tone was by itself, they didn't respond any more than usual. This supported the R-W theory: the Light "siphoned" all the predictive strength, leaving the tone with none.

Exclusive-OR / Negative Patterning Problem

This is a trickier problem for learning theories:

The rule: Signal A leads to food. Signal B leads to food. But when A and B are together (A+B), there is no food.
Example: Tone alone → Food; Click alone → Food; Tone + Click → No food.
What happens: Rats learn to respond strongly to the tone alone and the click alone, but they learn to stop responding when the tone and click are together, even though each one on its own means food.
Why it's a problem for standard R-W: Standard R-W says that Tone and Click gain strength when they're alone and lose strength when they're together. It predicts they should end up with zero strength overall, which doesn't match the fact that animals respond to them individually.

R-W Fix: “Configural Cue” or Element X

Scientists modified R-W to explain this. They added a new idea:

They suggested that when Tone and Click are presented together, the animal doesn't just see Tone and Click. It also sees a new, unique signal – let's call it "Stimulus X" – which represents the combination of Tone + Click.
How it works: When Tone and Click appear together (with no food), this special "Stimulus X" becomes strongly associated with no food (it develops a strong negative strength). Meanwhile, Tone and Click alone still gain positive strength (associated with food).
Outcome: So, when Tone + Click are together, their individual positive strengths are cancelled out by the strong negative strength of "Stimulus X." This explains why the animal doesn't respond to the compound, even though it responds to the individual cues. This modification made R-W able to explain configural learning (learning about combinations).

Conceptual & Practical Take-Homes

Animals can learn in different ways: sometimes they pay attention to individual signals, and sometimes they learn about the relationships between signals. It depends on the situation.
"Peak shift" shows how an inhibitory signal (S–) can affect where an animal responds most strongly, even when that signal isn't present.
Feature-positive experiments (like the light + tone) demonstrate why theories that assume signals share "credit" (like R-W) work better than those that treat each signal's strength purely individually.
Negative patterning shows that animals learn about combinations of signals, not just the individual parts. This led to more advanced learning models.

Connections & Implications

Modern machine-learning: Ideas like "configural cues" (Stimulus X) are similar to how modern computer programs (like "convolutional networks") learn to recognize complex patterns by looking at features formed by combining simpler ones.
Applied behavior analysis: When training detection dogs (e.g., for bombs), trainers use methods similar to feature-positive discrimination. They might only reinforce the dog for finding the target smell when a specific "search cue" is present, which helps the dog avoid false alarms.
Ethical/philosophical: The fact that animals can perform relational judgments suggests they have richer thinking abilities than just being simple "reflex machines." This changes how we view animal intelligence.

Key Terms & Symbols (Quick Reference)

$S+$ : The signal that is followed by something good (like food).
$S-$ : The signal that is not followed by something good.
Generalization Gradient: A graph showing how an animal responds less and less as a new signal becomes less similar to the original trained signal.
$\lambda$ (lambda): The maximum possible amount of "associative strength" or expectation for a given outcome (like the maximum expectation of getting food).
$V$ : The current amount of "associative strength" a signal has (how much the animal expects something to happen when that signal is present).
Configural Cue $X$ : A made-up signal that represents when two or more individual signals appear together, acting like a unique combined signal.
Feature-Positive: A type of learning task where an extra signal tells the animal that something good will happen.
Peak Shift: When the highest level of response to a signal shifts away from the original S+ in the opposite direction of the S–.

Numerical / Statistical Nuggets Mentioned

Hanson peak shift: Pigeons were trained with S+ at green light ( $550\,\ ext{nm }$ ) and S– at yellow-green ( $590\,\ ext{nm }$ ). Their strongest response was around $540\,\ ext{nm }$ (which is closer to blue-green, shifted away from the yellow-green S–).
Gonzalez tins: The original tins for primates might have been, for example, 13 cm (small), 17 cm (medium), and 21 cm (large). When tested, they picked the "new" intermediate size from a different set of tins.
Exclusive-OR task: Rats usually start to learn and show a strong difference in response (responding to individual cues but not the combination) after about 9 to 11 training sessions, with 60 trials in each session.