Fig. 1. Illustration of a classical conditioning experiment. When the fish sees the food falling down (unconditioned stimulus — UCS), it swims from the ground to the upper-left corner (unconditioned response — UCR) to eat it. If a light (conditioned stimulus — CS) flashes some seconds before the arrival of the UCS, the fish would start moving up (conditioned response — CR) right after the light is presented, since it predicts that the food is coming soon.

Fig. 2. Basic neural network architecture for the SB model with two CSs and a single UCS. A receptive layer (left) represents the stimuli, and the output layer (neuron to the right) the response. Learning takes place at the synaptic sites or connections between each stimulus and the response (small circles) according to Eq. (1).

Fig. 3. Three stages of stimulus conditioning. (a) Acquisition phase: the UCS onset follows the CS offset, yielding the UCR and an increase of the association between the CS and the response. (b) Prediction of the UCS: the CS generates a CR that is followed by the UCR elicited by the UCS, the response is correctly anticipated and the weight does not undergo any change. (c) Extinction phase: the CS generates a CR, but the UCS does not follow after the CS offset, which causes a decrease in the synaptic weight.

Fig. 4. The user interface. The user feeds a database with sentences of two or three words, chosen from the three menus that can be opened up by pressing the corresponding button (pronoun, verb or object, complement). The sentence is sent to the database and cleaned from the window when the ‘OK’ button is pressed. Since these list box selection menus are closed when a word is chosen, the user needs two clicks to select a word: first to open up the menu, and then to select the word.

Fig. 5. Two different sequences of events processed by the interface. User's commands are represented over the lines, with B1, B2 and B3 meaning to press a Pronoun, Verb or Complement button, respectively; and S1, S2 and S3 meaning to choose a word from the displayed menu. The actions performed by the interface are shown below the line, where m1, m2, m3 mean to open up the Pronoun, Verb or Complement menu, respectively, and highlighting the first word in the list; actions w1, w2 and w3 imply that the selected word (pronoun, verb or complement, respectively) is highlighted, then the word appears on the display and finally the menu is closed; action s erases the sentence from the display and sends it to a database.

Fig. 6. Neural network architecture. Big circles represent neurons. Small circles are synaptic connections between pairs of neurons each one consisting in a synaptic weight (W_ij) that represents the association level between the stimulus and the response. Solid synapses are nonmodifiable excitatory links, and empty synapses represent plastic connections (adjustable according to Eq. (1)). For simplicity, adjustable connections from some inputs to the output neurons are omitted. Pressing button B1 activates the corresponding neuron in layer 1 and this activity is transmitted to its counter-part in layer 2, generating action m1; selecting a pronoun (S1) activates its corresponding neuron in layer 1 and the activity is transmitted to its counter-part in layer 2, generating action w1, and so forth for the following menu. Learning takes place in adjustable synapses because a trace of the layer 1 activity is stored in each neuron and linked to the layer 2 activity.

Fig. 7. Weight modification for some classical conditioning phenomena. (a) Acquisition and extinction of a conditioned response. During trials 1–100, the CS was presented followed by a UCS, and alone from trials 101–200. The synaptic weight increases and saturates as the CS successfully predicts the UCR in the first experiment, and goes down in the second experiment, since the CR is not validated by the arrival of the UCS. (b) Conditioned inhibition. After an acquisition experiment (trials 1–80), CS1 and CS2 are presented together without the UCS and, alternatively, CS1 is presented alone preceding the UCS. Although initially the association between CS1 and the UCR started to decrease (solid line), the system realizes later that it is the CS2 that predicts the absence of the UCS, and a large inhibitory connection (dotted line) grows up to cancel the generation of the UCR by CS1.

Fig. 8. Cumulative time for (a) the first, and (b) second sequences of events in Fig. 5. The lines represent the delays in clicking on the buttons for beginners (dotted lines) and experts with (dashed lines) and without adaptability (solid lines). The value at B1 represents the delay since the ‘OK’ button was pressed, that is very small for the solid line, because the interface anticipates it.