An attention-modulated associative network

Abstract

We present an elemental model of associative learning that describes interactions between stimulus elements as a process of competitive normalization. Building on the assumptions laid out in Harris (2006), stimuli are represented as an array of elements that compete for attention according to the strength of their input. Elements form associations among each other according to temporal correlations in their activation but restricted by their connectivity. The model moves beyond its predecessor by specifying excitatory, inhibitory, and attention processes for each element in real time and describing their interaction as a form of suppressive gain control. Attention is formalized in this model as a network of mutually inhibitory units that moderate the activation of stimulus elements by controlling the level to which the elements are suppressed by their own inhibitory processes. The model is applied to a range of complex discriminations and related phenomena that have been taken as evidence for configural-learning processes.

Serial overshadowing of taste aversion learning by stimuli preceding the target taste

“…Unlike the trial-based formulation of the Rescorla-Wagner model, realtime models do tend to predict at least some one-trial overshadowing due to the manner in which a learning episode is broken into small time steps or microtrials, updating prediction error incrementally. However, this predicted one-trial effect is typically weaker relative to multiple-trial learning, especially when trace conditioning is simulated using models that explicitly assume some form of stimulus interference (e.g., Harris & Livesey, 2010). The clearest reason why a real-time error correction model would predict more onetrial overshadowing for the first of two serially presented stimuli than for the second is that the representation of the second stimulus disrupts the representation of the first.…”

“…For example, instead of the acquisitiondeficit account of overshadowing that is assumed here, it is possible that one based on retrieval deficits, such as a comparator theory (e.g., Denniston, Savastano, & Miller, 2001), could be modified to fit the present results. Furthermore, several real-time models based on learning mechanisms similar to the Rescorla-Wagner model can account for these data via processes similar to those we have outlined here (e.g., Harris & Livesey, 2010;McLaren & Mackintosh, 2000;Wagner, 1981). By their very nature, these models provide scope for quantitative differences in trace decay, although they make different assumptions about how stimulus representations might interfere with one another.…”

“…Thus, our definition departs from the conventional use of the term to denote the change caused purely by the passage of time. We chose an exponential decay function because it naturally follows from the formulae of many real-time models of associative learning (e.g., Harris & Livesey, 2010;McLaren & Mackintosh, 2000). However, in principle, similar results could be achieved with either linear or sigmoidal functions.…”

Comparing negative patterning and biconditional discrimination in a simulated foraging task

“…Under these conditions, positive patterning and biconditional discrimination were substantially easier than negative patterning, as predicted by configural cue theories like Pearce (1994) and Rescorla and Wagner (1972). This finding is problematic for models such as McLaren and Macintosh (2002) and Harris and Livesey (2010), which either or omit or minimize the role of configural elements. 5, 6, and 7 Pearce's (1994) configural model, the Rescorla and Wagner (1972) model, and Harris's (2006) model in which all stimuli are compounded with novel stimuli.…”

“…Harris, Livesey, Gharaei, & Westbrook, 2008;Harris & Livesey, 2010). In this model, elements compete directly for attention.…”