Repetition Suppression and Its Contextual Determinants in Predictive Coding
Total Page:16
File Type:pdf, Size:1020Kb
cortex xxx (2016) 1e16 Available online at www.sciencedirect.com ScienceDirect Journal homepage: www.elsevier.com/locate/cortex Special issue: Review Repetition suppression and its contextual determinants in predictive coding * Ryszard Auksztulewicz and Karl Friston Wellcome Trust Centre for Neuroimaging, Institute of Neurology, University College London, London, United Kingdom article info abstract Article history: This paper presents a review of theoretical and empirical work on repetition suppression in Received 8 June 2015 the context of predictive coding. Predictive coding is a neurobiologically plausible scheme Reviewed 27 August 2015 explaining how biological systems might perform perceptual inference and learning. From Revised 7 September 2015 this perspective, repetition suppression is a manifestation of minimising prediction error Accepted 11 November 2015 through adaptive changes in predictions about the content and precision of sensory inputs. Published online xxx Simulations of artificial neural hierarchies provide a principled way of understanding how repetition suppression e at different time scales e can be explained in terms of inference Keywords: and learning implemented under predictive coding. This formulation of repetition sup- Repetition suppression pression is supported by results of numerous empirical studies of repetition suppression Predictive coding and its contextual determinants. Mismatch negativity © 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC Perceptual inference BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Perceptual learning Egner, 2008). The predictive coding framework provides a 1. Introduction principled explanation of repetition effects in terms of perceptual inference and learning, mediated by changes in The effect of stimulus repetition on neural responses is one of synaptic efficacy (Friston, 2005). Adaptive changes in coupling the most studied phenomena in neuroscience. Typically, of neuronal populations within areas and connectivity be- repeated stimuli evoke neural activity with amplitudes tween areas are means of optimising a neuronal (generative) smaller than responses to novel stimuli. Although repetition model of the external world to provide more accurate and suppression is often portrayed as an expression of relatively precise predictions about sensory inputs. Thus, repetition simple mechanisms, such as neural fatigue (Grill-Spector, suppression can be understood in terms of ‘explaining away’ Henson, & Martin, 2006), its dependence on statistical regu- sensory prediction errors. larities in the environment and other contextual factors casts In the following, we will review modelling and experi- repetition suppression as a consequence of sensory pre- mental work on repetition suppression in the setting of pre- dictions (e.g., Summerfield, Trittschuh, Monti, Mesulam, & dictive coding. First, we introduce the predictive coding * Corresponding author. Wellcome Trust Centre for Neuroimaging, Institute of Neurology, University College London, 12 Queen Square, London WC1N 3BG, United Kingdom. E-mail addresses: [email protected] (R. Auksztulewicz), [email protected] (K. Friston). http://dx.doi.org/10.1016/j.cortex.2015.11.024 0010-9452/© 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http:// creativecommons.org/licenses/by-nc-nd/4.0/). Please cite this article in press as: Auksztulewicz, R., & Friston, K., Repetition suppression and its contextual determinants in predictive coding, Cortex (2016), http://dx.doi.org/10.1016/j.cortex.2015.11.024 2 cortex xxx (2016) 1e16 framework and portray neuronal message passing in terms of In the equations above, v denote causes representing descending predictions, ascending prediction errors, and (hidden) causes (e.g., the bark of a dog), while x denote states modulatory precision. We will then show how the predictive of the world mediating the influence of that cause on sensory coding scheme can be mapped onto a canonical cortical signals (e.g., the acoustic consequences of a dog barking). microcircuit. The subsequent section will focus on explaining Because these dynamics follow stereotyped trajectories over the dynamics of repetition suppression using simulations and time, they endow the model with memory. In equations computational modelling of empirical data. This will be fol- above, tilde is used to augment the variables with their lowed by a review of empirical studies on repetition sup- generalised coordinates of motion, i.e., pression and its context sensitivity, with a special focus on the  à ~ ¼ ; 0; 00 ; … : crucial role of predictions and precision in modulating the x x x x (3) effects of stimulus repetition. Eq. (1) describes the motion of states x(i) (at i-th hierarchical level) as a nonlinear function f of causes and states them- selves. Here D is a block-matrix derivative operator, with 2. Predictive coding identity matrices on its first leading-diagonal. Eq. (2) describes the motion of causes at a hierarchically lower level i-1 as a In order to maintain their integrity (e.g., homoeostasis), bio- nonlinear function g of hidden causes and states at the level logical systems have to minimise the excursions or entropy of above. Random fluctuations in hidden causes and states are u~ðiÞ u~ðiÞ their interoceptive and exteroceptive states. Since entropy is denoted by v and x respectively. the average of surprise (also known as surprisal or self- Since the brain does not have direct access to the causes information) over time, biological systems should continu- and states in the external world, it can only infer the most ally minimise their surprise about sensory states. Mathe- likely values under its generative model: mathematically, matically, this is equivalent to maximising the Bayesian these values are expectations. In other words, the generative evidence for their model of sensory inputs, also known as model maps from causes to sensory consequences, while Bayesian filtering. Predictive coding (Friston, 2005; Mumford, perception solves the (usually very difficult) inverse problem 1992; Rao & Ballard, 1999) is a popular, neurobiologically which is to map from sensations to their underlying causes. plausible Bayesian filtering scheme that decomposes the An inversion of hierarchical dynamic models can be cast in optimisation of the agent's (neuronal) model of the world into terms of a hierarchical message passing scheme also known two tractable components; namely (1) changes in expecta- as predictive coding: tions about the sensory inputs and (2) the computation of ð Þ m~_ i ¼ m~ðiÞ À v ~εðiÞxðiÞ À xðiþ1Þ prediction errors that are needed to change expectations. v D v v~ v (4) Minimising surprise e or maximising model evidence e ð Þ lies at the heart of the free energy principle, where free energy ~_ i ~ðiÞ ðiÞ ðiÞ m ¼ Dm À v~~ε x (5) provides a proxy for surprise that, under simplifying as- x x x sumptions, can be reduced to prediction error (Friston, 2010). Á xðiÞ ¼ PðiÞ~εðiÞ ¼ PðiÞðm~ðiÀ1Þ À ðiÞðm~ðiÞ; m~ðiÞ This means one can understand the process of perception as v v v v v g x v (6) the resolution of prediction errors, by changing top-down xðiÞ ¼ PðiÞ~εðiÞ ¼ PðiÞð m~ðiÞ À ðiÞðm~ðiÞ; m~ðiÞÞ predictions about the causes of sensory input (Fig. 1). Intui- x x x x D x f x v (7) tively, the predictions descending along the processing (e.g., cortical) hierarchy are compared against sampled sensory This message passing suggests two distinct populations of neurons: one encoding the trajectory of the expectations inputs in sensory cortex (or expectations as intermediate hi- ð Þ ð Þ m~_ i m~_ i erarchical levels). The ensuing prediction errors are then (conditional means) of hidden causes v and states x , passed up the hierarchy to optimise expectations and subse- which we can therefore label state-units, and one encoding ~εðiÞ ~εðiÞ quent predictions. If the ascending input matches the the prediction errors v and x weighted by their respective PðiÞ PðiÞ descending prediction, the prediction error will be low e as precisions v and x , which we can label error-units. exemplified by repetition suppression. If the predictions are These precisions are the inverse amplitude of the random inconsistent with the incoming input, the prediction error will fluctuations above, so that when the fluctuations are be high e as illustrated by mismatch negativity (Garrido, small, prediction errors become precise and are amplified. v v Kilner, Stephan, & Friston, 2009). In the following, the notion To simplify notation, x~ and v~ are used to denote a of perception under predictive coding will be unpacked in the partial derivative with respect to hidden states and causes v context of repetition suppression. respectively. Temporal derivatives, e.g., tx,aredenotedbya _ The ability of the brain to infer the causes of its sensations dot x. rests upon the presence of statistical structure or contin- These equations may look complicated but are formally gencies in the environment. These contingencies can be simple and quite revealing in terms of which (neuronal) units embodied within a generative model describing the hierar- talk to each other. In brief, the equations suggest that error- chical and dynamic statistics of the external world: units receive messages from populations in the same hierar- À Á chical level and the level above, while state-units are driven by ~ðiÞ ¼ ðiÞ ~ðiÞ; ~ðiÞ þ u~ ðiÞ Dx f x v x (1) error-units in the same level and the level below. The pre- ð þ Þ ð Þ diction errors from the same level x i 1 and the level below x i À Á ð À Þ ð Þ ð Þ ð Þ v~ i 1 ¼ gðiÞ x~ i ; v~ i þ u~ i : provide lateral and bottom-up messages driving the condi- v (2) ð Þ tional expectations m~ i towards better lateral and top-down Please cite this article in press as: Auksztulewicz, R., & Friston, K., Repetition suppression and its contextual determinants in predictive coding, Cortex (2016), http://dx.doi.org/10.1016/j.cortex.2015.11.024 cortex xxx (2016) 1e16 3 Fig.