Let us modify the RW-model to account for varying CS salience, as well as to include a putative discrimination threshold in the following equation:
Where α(t) represents variable salience over time and α
min is the salience threshold for learning to occur. For simplicity, we represent λ as a sliding logistic function of α , because the quality of sensory representation should degrade gradually as salience reaches α
min, compromising discrimination  and learning. We assume that discrimination performance is constrained by a perceptual grid that filters out relevant information for the discrimination task, as represented by λ(α) at low α values.
However, λ could also be modeled using a Boltzmann distribution , or other functions [13, 14]. (Note that additional variants on the model have been addressed elsewhere [7, 15]).
Regarding varying salience: if stimulus 'i' is reinforced, then α
(t) should increase, and if stimulus 'j' is not reinforced, then α
(t) should decrease. In a situation where the stimuli, 'i', 'j', and 'k' are sequentially reinforced, then an increase in a α
(t) should affect α
(t) and α
(t) according to the degree of similarity between the stimuli. Therefore, the varying salience over time may adopt the following form:
where Si,j represents the degree of similarity between the ith (reference) and jth stimuli (0 ≤ S ≤ 1), and α
(t) is the dynamic representation of salience with respect to item 'i', as the probability of attention will vary together with salience and learning [16, 17]. Thus, α (t) should increase or decrease depending on both, reinforcement levels and the temporal arrangement of stimuli similarity during training. Evidently, we do not know how salience evolves with learning. Let us consider a simple steady-state scenario, where α (t) equals Si,j. What would be the effect of varying stimuli similarity during learning? To explore this idea, we first generated a set ofstimuli with different degrees of similarity by using random numbers from normal distributions with fixed meanand variable standard deviations (Figure 1A). These numbers represent training stimuli with different salience. To investigate whether variable salience has a relevant effect in learning, we sorted the stimuli using other decreasing (black line) or increasing (gray line) similarity (Figure 1B). These arrangements maximize the relative difference in salience between training programs but consist of exactly the same stimuli. Next, we calculated λ(α), applying either no salience threshold (i.e. α
min = 0) or a putative threshold of 0.3 (α
min = 0.3; Figure 1C). Panels D-E show the predicted learning curves, as given by Eq.2. In all cases of identical mean salience of 0.5, the temporal arrangement of training stimuli determined the shape of the learning curves.
Moreover, when discriminative training involved stimuli below the salience threshold for learning (Figure 1E), stimuli with salience below α
min were undetectable, V(t) did not increase (for V(t) = 0), and the curves decayed in a mono-exponential manner due to the lack of reinforcement (0 ≤ V(t) ≤ 1). When similarity was held constant (thick dotted lines), the learning curves were identical to those predicted by the standard model.