James, M. 1, 2 , Li, M. 1, 2 , You, H. 1, 2 , Luthra, S. 1, 2 & Steiner, R. 1, 2
1 Psychological Sciences, U. of Connecticut
2 Connecticut Institute for the Brain and Cognitive Sciences
Predictive processing (PP) and predictive coding (PC) are sometimes used synonymously. However, maintaining a distinction may be useful theoretically. Formally, PC is a top-down Bayesian architecture that emphasizes novel information in the form of prediction error (Rao & Ballard, 1999), whereas PP refers to any neural or behavioral indicator of prediction or anticipation. Cognitive neuroscientists often interpret PP as indicating PC, sometimes in cases that may not be prima facie relatable to the same computational mechanism. For example, Gagnepain et al. (2012) interpreted lesser MEG responses for expectation-consistent speech as PC (and linked their results to a formal model), while Summerfield et al. (2006) interpreted stronger BOLD responses for expected faces as PC. This raises the question of how neural and behavioral data can be linked to computational models. We compare a recent PC model with two older models: one that may seem intuitively inconsistent with PP and PC (the interactive activation model, TRACE [McClelland & Elman, 1986]) and one that seems intuitively consistent with PP and PC (the Simple Recurrent Network [SRN; Elman, 1990]). We consider how each may be linked to PP and PC. In particular, we examine how competitive and cooperative weights and activation flow in TRACE and the SRN lead to clear PP in both models and how each might relate to PC. We also build on recent work by McClelland (2013) and McClelland et al. (2014) to assess the generative models embodied in interactive activation and recurrent network models, and compare them to explicit models of PC.