Generalization gradient

The generalization gradient describes a graphical function which, in the context of discrimination learning, indicates how much an organism's reactions from a certain stimulus to other stimuli generalize.

Learning to discriminate as a prerequisite

Before a generalization gradient can be determined experimentally, the organism must go through a discrimination learning phase. Discrimination learning takes place in both classical and operant conditioning. Discrimination learning is explained in the following in the form of instrumental and operant conditioning , whereby the organism learns to show the instrumental reaction only when certain stimuli are presented.

Since discrimination experiments are carried out in learning psychological practice with animals (rats, mice, pigeons, etc.), it should be illustrated using an example with pigeons and visual stimuli that has been well proven in research:

Experimental set-up

The pigeon is in a Skinner box during the instrumental learning phase . This is a small cage, which in our example is equipped with a small feed dispenser in which feed pellets can be presented. There is also a small illuminated button and another light emitting diode or lightbulb in front of the pigeon.

Simple operant conditioning

With "simple" operant conditioning, the pigeon could be reinforced to peck the little illuminated button. As soon as it shows a picking reaction, an amplifier - a feed pellet - is presented in the feed dispenser. Over time, the pigeon learns to show the pecking reaction very often and continuously in order to receive the amplifier. How often the pigeon has to show the reaction or how much time has to pass after the reaction until a reinforcer becomes available is determined by the defined reinforcement plan . Here a rough distinction is made between ratio plans (here the number of reactions until the reinforcement is given is decisive) and interval plans (here the reinforcer becomes available after a certain period of time after a reaction).

Types of Discrimination Learning

Discrimination learning is a form of stimulus control of behavior. If there is behavior and stimulus control, this means that the organism shows changes in its behavior when the stimuli change. Learning to discriminate can also very often be observed in reality, e.g. B. Do we behave differently in the presence of our friends than in the presence of a boss. The people present thus become discriminative stimuli.

In the experiment with our pigeon we want to use the second little light as a discriminative simulus.

S + learning

The simplest form of learning to discriminate is the introduction of a so-called S + (also called SD or S *). When this stimulus is presented, the instrumental response is amplified - in the absence of the stimulus it is not amplified. With continuous training, the organism learns to show the instrumental reaction only when the S + is present.

In the example with the dove, we could introduce the second little light as S +. Whenever it is lit, the pecking on the button is intensified. When the light is out, the pecking is not amplified.

S− learning

Another form of learning to discriminate is the introduction of an S- (also called S-delta). When a stimulus is presented, the response is never amplified, and it is only amplified in the absence of the stimulus. After sufficient training, the organism learns to show the reaction only in the absence of the stimulus.

In the pigeon example we would only intensify the pecking when the light is out. Presentation of the light signals the lack of amplification.

S + / S− learning

You can now also combine S + and S - learning. In the presentation of a certain stimulus (S +) the reaction is intensified. When a different stimulus is presented (S−), the reaction is not intensified.

A classic example of this is the use of different wavelengths of light in the pigeon experiment. Roughly speaking, different wavelengths mean different colors of light. So we could illuminate the light green as S + and show a red light as S-. The picking reaction is thus intensified during the green light and never intensified during the red light. Over time the pigeon learns to show the reaction strongly when the green light is presented and hardly when the red light is presented.

Likewise, we could have introduced a sound as S + and a light as S−. If, however, as in the example with the different wavelengths of light, a distinction is made between S + and S− on one dimension (here: wavelength), one speaks of intradimensional discrimination . The opposite is the use of two stimuli from different dimensions - e.g. B. Light and Sound - represent.

Generalization gradients

detection

Generalization means that the organism reacts to two different stimuli with similar or identical behavior (i.e. exactly the opposite of stimulus control of behavior). As already mentioned, a discrimination learning phase must have preceded the determination of the generalization gradient. Now you do the following: In a test phase you offer different stimuli one after the other, which differ more or less from the originally trained discriminative stimulus. Now you measure the frequency of the instrumental reaction during the different stimuli. If the behavior of the organism has come very strongly under stimulus control, one hardly observes generalization - ie deviating test stimuli cause different behavior. However, if the behavior is hardly subject to stimulus control, strong generalization is observed - that is, the behavior is almost unchanged with different test stimuli. It is of course important that the organism can adequately perceive the test stimuli at all (so a pigeon can no longer perceive light above a certain wavelength). In addition, the test stimuli are usually varied intradimensionally by the training stimulus. This means that only the strength of one characteristic expression is changed (e.g. light of different wavelengths as test stimuli).

A generalization gradient is then drawn in a two-dimensional coordinate system. The dimension in which the test stimuli are varied (e.g. the wavelength of the light) is plotted on the x-axis. On the y-axis, the frequency of the instrumental response.

Types of generalization gradients

The shape of the generalization gradient depends on the type of previous discrimination learning.

Excitatory generalization gradient

After learning to discriminate with an S + one obtains an excitatory gradient. This has its peak ( peak ), so the highest response rate to the original S + and falls on both sides with increasing dissimilarity of the test stimuli to the S + like a mountain from. The steeper the gradient, the more the behavior is under stimulus control.

Inhibitory generalization gradient

An inhibitory gradient is obtained after S-learning. This has a similar structure to the excitatory gradient, only that it is mirrored horizontally. It has its lowest point (i.e. the lowest reaction frequency) at the original S− and increases with increasing dissimilarity of the test stimuli with the S−. It thus has the shape of a pointed valley with the lowest point above the S−.

S + / S− generalization gradient

After intradimensional S + / S- learning, another form of gradient is obtained, which has some peculiarities. It is important that S + and S− only differ in one dimension (e.g. the sound frequency or the wavelength of light). During the test phase, various test stimuli are offered that differ on the same dimension as S + and S−. The resulting gradient has a shape similar to the excitatory gradient (i.e. a "mountain shape") with its peak at S + and the lowest frequency at S−. However, if S + and S− are very close to one another within the dimension (e.g. S + is light with 500 nm and S− light with 520 nm wavelength), one observes some peculiarities:

Peak shift

If S + and S− are intradimensionally very close to one another, the peak (i.e. the tip) of the gradient is no longer exactly above the S +, but slightly shifted away from it. The shift takes place in the direction in which the S− is not .

If the S + z. B. light with 500 nm and the S− light with 520 nm wavelength, then one observes that the peak is now perhaps 480 nm instead of 500 nm (S +).

Behavioral Contrast

There is the long-studied theory that the generalization gradient after S + / S− learning corresponds to the net gradient from an S + and an S− gradient. In practice, this assumption does not seem unfounded, but after S + / S− learning one observes a steeper (i.e. faster falling) and higher gradient (i.e. with a higher peak) than one would predict from the combination of two S + and S− gradients. This phenomenon is called behavioral contrast.

Individual evidence

↑ Christian Becker-Carus, Mike Wendt: General Psychology: An Introduction . Springer-Verlag, 2017, ISBN 978-3-662-53006-1 , pp. 320 ( limited preview in Google Book search).
↑ Katja Mackowiak, Gerhard W. Lauth, Ralf Spieß: Promotion of learning processes . Kohlhammer Verlag, 2008, ISBN 978-3-17-028064-9 ( google.de [accessed on February 19, 2019]).

[1] Christian Becker-Carus, Mike Wendt: General Psychology: An Introduction . Springer-Verlag, 2017, ISBN 978-3-662-53006-1 , pp. 320 ( limited preview in Google Book search).

[2] Katja Mackowiak, Gerhard W. Lauth, Ralf Spieß: Promotion of learning processes . Kohlhammer Verlag, 2008, ISBN 978-3-17-028064-9 ( google.de [accessed on February 19, 2019]).