Uncertainty in perception and the Hierarchical Gaussian Filter.

Mathys CD, Lomakina EI, Daunizeau J, Iglesias S, Brodersen KH, Friston KJ, Stephan KE. · 2014

Read Uncertainty in perception and the Hierarchical Gaussian Filter. by Mathys CD, Lomakina EI, Daunizeau J, Iglesias S, Brodersen KH, Friston KJ, Stephan KE. (2014) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc4237059.txt.

Introduction

Christoph D. Mathys, Klaas E. Stephan, and five co-authors generalized a brain computation model in Frontiers in Human Neuroscience in 2014. Perception, in the Bayesian view, is the brain guessing the causes of its senses while tracking how uncertain those guesses are. The Hierarchical Gaussian Filter, or HGF, turns that idea into equations that can be fitted to a person's behaviour.

The paper extends the model's hierarchy to any number of levels and shows how different kinds of uncertainty enter its update rules. It also pairs the filter with decision models so choices, not just beliefs, can be predicted. Simulations compared four fitting methods across noise levels; all coped well, and variational Bayes gave the best mix of speed and information. The work aims to make the model a standard tool for studying perception gone wrong.

Catalog record

CREATORMathys CD, Lomakina EI, Daunizeau J, Iglesias S, Brodersen KH, Friston KJ, Stephan KE.
YEAR2014
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC4237059

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Uncertainty in perception and the Hierarchical Gaussian Filter By Christoph D Mathys, Ekaterina I Lomakina, Jean Daunizeau, Sandra Iglesias, Kay H Brodersen, Karl J Friston, Klaas E Stephan Published in: Frontiers in Human Neuroscience DOI: 10.3389/fnhum.2014.00825 ABSTRACT Abstract In its full sense, perception rests on an agent's model of how its sensory input comes about and the inferences it draws based on this model. These inferences are necessarily uncertain. Here, we illustrate how the Hierarchical Gaussian Filter (HGF) offers a principled and generic way to deal with the several forms that uncertainty in perception takes. The HGF is a recent derivation of one-step update equations from Bayesian principles that rests on a hierarchical generative model of the environment and its (in)stability. It is computationally highly efficient, allows for online estimates of hidden states, and has found numerous applications to experimental data from human subjects. In this paper, we generalize previous descriptions of the HGF and its account of perceptual uncertainty. First, we explicitly formulate the extension of the HGF's hierarchy to any number of levels; second, we discuss how various forms of uncertainty are accommodated by the minimization of variational free energy as encoded in the update equations; third, we combine the HGF with decision models and demonstrate the inversion of this combination; finally, we report a simulation study that compared four optimization methods for inverting the HGF/decision model combination at different noise levels. These four methods (Nelder–Mead simplex algorithm, Gaussian process-based global optimization, variational Bayes and Markov chain Monte Carlo sampling) all performed well even under considerable noise, with variational Bayes offering the best combination of efficiency and informativeness of inference. Our results demonstrate that the HGF provides a principled, flexible, and efficient—but at the same time intuitive—framework for the resolution of perceptual uncertainty in behaving agents. Keywords: uncertainty, volatility, Bayesian inference, hierarchical modeling, filtering, free energy, learning, decision-making INTRODUCTION Perception has long been proposed to take place in the context of prediction (Helmholtz, 1860). This entails that agents have a model of the environment which generates their sensory input. Probability theory formally prescribes how agents should learn about their environment from sensory information, given a model. This rests on sequential updating of beliefs according to Bayes' theorem, where beliefs represent inferences about hidden states of the environment in the form of posterior probability distributions. It is this process that we refer to as perception. Beliefs about hidden states are inherently uncertain. This uncertainty has two sources. First, even when states are constant, the amount of sensory information will in general be too little to infer them exactly. This has been referred to as informational uncertainty or estimation uncertainty (Payzan-LeNestour and Bossaerts, 2011). The second source of uncertainty is the possibility that states change with time, i.e., environmental uncertainty. Various models have suggested how an agent may deal with an environment fraught with both kinds of uncertainty (e.g., Yu and Dayan, 2003, 2005; Nassar et al., 2010; Payzan-LeNestour and Bossaerts, 2011; Wilson et al., 2013). Here, we discuss an alternative approach that derives closed form update equations for the hidden states, and crucially, for the uncertainty about them, by variational inversion of a generic hierarchical generative model that reflects the time-varying structure of the environment in its higher levels. This derivation has the advantage that the resulting updates optimize a clearly defined objective function, namely variational free energy. Since this quantity is an approximation to surprise (i.e., to the negative log-probability of sensory input), the updates are optimal in the sense that they minimize surprise, given an agent's individual model. Furthermore, the updates explicitly reflect informational and environmental uncertainty. Our approach makes use of a framework which assumes that agents have an internal generative model of their sensory input. This model is generative in the sense that it describes how sensory inputs are generated by the external world. It does this by assigning a probability (the likelihood) to each sensory input given states (which vary with time) and parameters (which are constant in time) and by completing this with a prior probability distribution for states and parameters. While the purpose of the model is to predict input emanating from the external world, it is internal in the sense that it reflects the agent's beliefs about how sensory inputs are generated by the external world. While Bayesian belief updating is optimal from the perspective of probability theory, it requires computing complicated integrals which are not tractable analytically and difficult to evaluate in real time. Although some attempts to design a Bayesian model of how biological agents learn in a changing environment were remarkably successful in explaining empirical behavior (Behrens et al., 2007, 2008), they were restricted by the computational burden imposed by these models and the assumption that the learning process was identical across subjects. Recently, however, theoretical advances have enabled computationally efficient approximations to exact Bayesian inference during learning (e.g., Friston, 2009; Daunizeau et al., 2010a,b) and have furnished a basis for biologically plausible mechanisms that might underlie belief updating in the brain. These approaches rest on variational Bayesian techniques which optimize a free-energy bound on the surprise about sensory inputs, given a model of the environment, and represent a special case of the general “Bayesian brain” hypothesis (Dayan et al., 1995; Knill and Pouget, 2004; Körding and Wolpert, 2006; Friston, 2009; Doya et al., 2011). This hypothesis has been highly influential in recent years, shaping concepts of brain function and inspiring the design of many specific computational models (see Friston and Dolan, 2010, for review). However, for practical applications to empirical data, a general purpose modeling framework has been lacking that would allow for straightforward “off the shelf” implementations of models explaining trial-wise empirical data (e.g., behavioral responses, eye movements, evoked response amplitude in EEG etc.) from the Bayesian brain perspective. This is in contrast to reinforcement learning (RL) models which, due to their simplicity and computational efficiency, have found widespread application in experimental neuroscience, for example, in the analysis of functional magnetic resonance imaging (fMRI) and behavioral data (for reviews, see Daw and Doya, 2006; O'Doherty et al., 2007). To fill this gap and provide a generic, robust and flexible framework for analysis of trial-wise data from the Bayesian brain perspective, we recently introduced the Hierarchical Gaussian Filter (HGF), a hierarchical Bayesian model Mathys et al. (2011) in which states evolve as coupled Gaussian random walks, such that each state determines the step size of the evolution of the next lower state (for examples of applications, cf. Iglesias et al., 2013; Joffily and Coricelli, 2013; Vossel et al., 2013). Based on a mean field approximation to the full Bayesian solution, we derived analytic update equations whose form resembles RL updates, with dynamic learning rates and precision-weighted prediction errors. These highly efficient update equations made our approach well suited for filtering purposes, i.e., predicting the value of (and, crucially, the uncertainty about) a hidden and moving quantity based on all information acquired up to a certain point. Our original formulation (Mathys et al., 2011) only contained three levels; here, we extend the HGF explicitly to any number of levels and show that the update equations maintain the same form across all levels because they are derived on the

Read the complete text