In fact, PEs are reported in a bewildering breadth of (hundreds of) studies. These models describe how we learn where the bad things lurk and the good things live, and which actions to undertake to avoid them or seek them out respectively ( Schultz et al., 1997 Schultz and Dickinson, 2000 Wise, 2004). Predictions and PEs are also key concepts in models of reward learning, motivational control, and decision-making (e.g., Rescorla and Wagner, 1972 Pearce and Hall, 1980 Behrens et al., 2007 Niv and Schoenbaum, 2008). In these models, predictions of impending perceptual events help us quickly interpret and disambiguate noisy and ambiguous input ( Kersten and Yuille, 2003 Sterzer et al., 2008). Prediction errors (PEs) are prominent in models of perception ( Rao and Ballard, 1999 Lee and Mumford, 2003), which propose how prior expectations help us to make sense of our environments. The study of prediction and prediction error signals in the brain is encountered in the largely segregated research fields of motivational control and perception. When such a mismatch between prior expectations and reality arises, this is referred to as a prediction error. Since the world is a continuously changing and stochastic environment, these predictions likewise need to be continuously changed and fine-tuned on the basis of novel information, which may conflict with the organism’s prior expectations. In support of these theories, there are many striking examples of the predictive nature of neural computations, in visual ( Bar et al., 2006 Alink et al., 2010 Meyer and Olson, 2011), auditory ( Ulanovsky et al., 2003 Baldeweg, 2006 Todorovic et al., 2011), and somatosensory perception ( Akatsuka et al., 2007 van Ede et al., 2011), as well as in action ( Blakemore et al., 1998 Bestmann et al., 2008 Franklin and Wolpert, 2011), language ( Kutas and Hillyard, 1980), memory ( Erickson and Desimone, 1999 Kumaran and Maguire, 2006, 2009), cognitive control ( Alexander and Brown, 2011), and motivational value processing ( Schultz, 1998 Hare et al., 2008 Daw et al., 2011). These theories propose that the brain’s primary objective is to infer the causes of its sensory input by reducing surprise, in order to allow it to successfully predict and interact with the world. Recent theories have gone as far as surmising that the primary function of the neocortex may lie in the prediction of future states of the environment ( Hawkins, 2004 Friston, 2005 Bar, 2009). Our ability to perceive structure and predict future states in the world is a remarkable feat that evolution has bestowed upon us.

We suggest that while encoding of PE is a common computation across brain regions, the content and function of these error signals can be very different and are determined by the afferent and efferent connections within the neural circuitry in which they arise. We discuss where different types of PE are encoded, how they are generated, and the different functional roles they fulfill. Here, we will make an attempt to see the forest for the trees and consider the commonalities and differences of reported PE signals in light of recent suggestions that the computation of PE forms a fundamental mode of brain function. Prediction errors (PE) are a central notion in theoretical models of reinforcement learning, perceptual inference, decision-making and cognition, and prediction error signals have been reported across a wide range of brain regions and experimental paradigms. 2Center for Neural Science, New York University, New York, NY, USA.1Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen, Nijmegen, Netherlands.den Ouden 1,2*, Peter Kok 1 and Floris P.

Tales of arise wonder text