Learned predictiveness models predict opposite attention biases in the inverse base-rate effect.

Abstract

Several attention-based models of associative learning are built upon the learned predictiveness principle, whereby learning is optimised by attending to the most predictive features and ignoring the least predictive features. Despite their functional similarity, these models differ in their formal mechanisms, and thus may produce very different predictions in some circumstances. As we demonstrate, this is particularly evident in the inverse base-rate effect. Using simulations with a modified Mackintosh model and the EXIT model, we found that models based on the learned predictiveness principle can account for rare-outcome choice biases associated with the inverse baserate effect, despite making opposite predictions for relative attention to rare versus common predictors. The models also make different predictions regarding changes in attention across training, and effects of context associations on attention to cues. Using a human causal learning task, we replicated the inverse base-rate effect and a recently reported reduction in this effect when the context is not predictive of the common outcome, and used eye-tracking to test model predictions about changes in attention both prior to making a decision, and during feedback. The results support the predictions made by EXIT, where the rare predictor commands greater attention than the common predictor throughout training. In addition, patterns of attention prior to making a decision differed to those during feedback, where effects of using a partially predictive context were evident only prior to making a prediction.

Attention biases in the inverse base-rate effect persist into new learning

“…Patterns of gaze biases support the idea that greater relative attention is paid to Cue C on AC trials than to Cue B on AB trials, under typical base-rate designs (e.g., the inverse base-rate effect: Don et al, 2019; and the highlighting effect 2 : Kruschke et al, 2005). Don et al (2019) measured gaze biases to cues both prior to making a prediction and during feedback, and assessed how gaze patterns differed based on the global base rates of the outcomes. Manipulation of the global base rates should affect the associations between the context and the prevailing common outcomes, where the context comprises the incidental cues related to features of the experimental trials and participating in the experiment more generally.…”

“…This account has been formalised in Kruschke’s (2001b) extended ADIT (EXIT) model, which is based on learned predictiveness principles like those proposed by Mackintosh (1975). Yet, the EXIT model and variants of Mackintosh’s model have been shown to make different predictions regarding attention to cues in the inverse base-rate effect (Don et al, 2019). Although EXIT is a relatively complex model containing several mechanisms, Paskewitz and Jones (2020) have shown that the EXIT model only requires rapid attentional shifts or attentional competition components to explain most experimental effects.…”

“…We will describe the design of Don et al (2019; see Table 1) in detail because it is highly relevant to the current study. In our standard condition, one outcome was always paired with common compounds, while another was always paired with rare compounds, such that overall, the context will be more strongly associated with the common outcome.…”

“…This was compared with a balanced outcome condition, where each outcome was paired with one common compound, and one rare compound, such that each outcome was experienced equally across the course of the experiment, and the context will not be strongly associated with either outcome. This condition has been shown to reduce the strength of the inverse base-rate effect (Don et al, 2019; Don & Livesey, 2017). In the standard condition, we found gaze biases towards C on AC trials both prior to making a prediction and during feedback.…”

Can We Set Aside Previous Experience in a Familiar Causal Scenario?

“…Finally, participants are asked to judge which foods are causing Mr. X to suffer an allergic reaction, and this judgment may also take the form of a probability or cued recall judgment. The food allergist task has been used to study cue competition effects such as blocking and overshadowing (e.g., Shanks and Lopez, 1996 ; Aitken et al, 2000 ; Lovibond et al, 2003 ; Beckers et al, 2005a ; Mitchell et al, 2005 , 2006 ; Vandorpe et al, 2007 ; Livesey et al, 2013 , 2019b ; Luque et al, 2013 ; Uengoer et al, 2013 ), learning of preventative relationships such as in the case of conditioned inhibition ( Karazinov and Boakes, 2004 , 2007 ), complex rule learning tasks such as the patterning task ( Shanks and Darby, 1998 ; Wills et al, 2011 ; Don et al, 2020 ), as well as a host of phenomena related to learned attentional changes including the learned predictiveness effect ( Le Pelley and McLaren, 2003 ; Don and Livesey, 2015 ; Shone et al, 2015 ), the inverse base-rate effect ( Don et al, 2019 ), outcome predictability effects ( Griffiths et al, 2015 ; Thorwart et al, 2017 ), and other related transfer effects ( Livesey et al, 2019a ). Food allergies are relatively commonplace such that, by the time they enter the laboratory, participants have a lifetime of experience with food and its ability to cause allergic reactions in oneself or others.…”