Computational Neuropsychology and Bayesian Inference

Abstract

Computational theories of brain function have become very influential in neuroscience. They have facilitated the growth of formal approaches to disease, particularly in psychiatric research. In this paper, we provide a narrative review of the body of computational research addressing neuropsychological syndromes, and focus on those that employ Bayesian frameworks. Bayesian approaches to understanding brain function formulate perception and action as inferential processes. These inferences combine 'prior' beliefs with a generative (predictive) model to explain the causes of sensations. Under this view, neuropsychological deficits can be thought of as false inferences that arise due to aberrant prior beliefs (that are poor fits to the real world). This draws upon the notion of a Bayes optimal pathology - optimal inference with suboptimal priors - and provides a means for computational phenotyping. In principle, any given neuropsychological disorder could be characterized by the set of prior beliefs that would make a patient's behavior appear Bayes optimal. We start with an overview of some key theoretical constructs and use these to motivate a form of computational neuropsychology that relates anatomical structures in the brain to the computations they perform. Throughout, we draw upon computational accounts of neuropsychological syndromes. These are selected to emphasize the key features of a Bayesian approach, and the possible types of pathological prior that may be present. They range from visual neglect through hallucinations to autism. Through these illustrative examples, we review the use of Bayesian approaches to understand the link between biology and computation that is at the heart of neuropsychology.

Keywords: active inference; computational phenotyping; neuropsychology; precision; predictive coding.

FIGURE 1

Generative models. These schematics graphically illustrate the structure of generative models. (A) The simplest model that permits Bayesian inference involves a hidden state, s, that is equipped with a prior P(s). This hidden state generates observable data, o, through a process defined by the likelihood P(o|s) (vertical arrow). (B) It is possible to equip such a model with dynamically changing hidden states. To do so, we must specify the probabilities of transitioning between states P(s_τ+1|s_τ) (horizontal arrows). (C) Transitions between states may be influenced by the course of action, π, that is pursued. (D) Hierarchical levels can be added to the generative model (Friston et al., 2017d). This means that the processes that generate the hidden states can themselves be accommodated in the inferences performed using the model.

FIGURE 2

Hierarchy in the cortex. This schematic illustrates two key features of cortical organization. The first is hierarchy, as defined by laminar specific projections. Projections from primary sensory areas, such as area VI, to higher cortical areas typically arise from layer III of a cortical column, and target layer IV. These ascending connections are shown in red. In contrast, descending connections (in blue) originate in deep layers of the cortex and project to both superficial and deep laminae. The second feature illustrated here is the separation of visual processing into two, dorsal and ventral, streams. In terms of the functional anatomy implied by generative models in the brain, this segregation implies a factorization of beliefs about the location and identity of a visual object (i.e., knowing what an object is does not tell you where it is – and vice versa).

FIGURE 3

Dorsal and ventral streams. Here we depict a plausible mapping of simple generative models to the dual streams of the language (Left) and attention (Right) networks. We highlight the likelihood mappings that correspond to white matter tracts implicated in disconnection syndromes. The number 1 in the blue circle on the left highlights the mapping from the left temporoparietal region, which responds to spoken words (Howard et al., 1992), to the inferior frontal gyrus, involved in the dorsal articulatory stream (Hickok, 2012b). This region is well placed to deal with proprioceptive data from the laryngeal and pharyngeal muscles (Simonyan and Horwitz, 2011). The connection corresponds to the arcuate fasciculus and lesions give rise to conduction aphasia. The number 2 indicates the mapping from dorsal frontal regions that represent eye fixation locations to ventral regions associated with target detection and identity. This corresponds to the second branch of the superior longitudinal fasciculus. Lesions to this structure are implicated in visual neglect (Doricchi and Tomaiuolo, 2003; Thiebaut de Schotten et al., 2005).

FIGURE 4

The anatomy of precision. The ascending neuromodulatory systems carrying dopaminergic, cholinergic, and noradrenergic signals are shown (in a simplified form). Dopaminergic neurons have their cell-bodies in the ventral tegmental area (VTA) and the substantia nigra pars compacta (SNc) – two nuclei in the midbrain. The medial forebrain bundle contains the axons of these cells, and allows them to target neurons in the prefrontal cortex and the medium spiny neurons of the striatum. The nucleus basalis of Meynert is found in the basal forebrain. This is the source of cholinergic projections to the cortex (Eckenstein et al., 1988). Axons originating here join the cingulum. Neurons in the locus coeruleus project from the brainstem, through the dorsal noradrenergic bundle, and also join the cingulum to supply the cortex with noradrenaline (Berridge and Waterhouse, 2003).

FIGURE 5

The anatomy of visual neglect. Three lesions implicated in visual neglect are highlighted here. 1 – Disconnection of the second branch of the right superior longitudinal fasciculus (a white matter tract that connects dorsal frontal with ventral parietal regions); 2 – Unilateral lesion to the right putamen; 3 – Unilateral lesion to the right pulvinar (a thalamic nucleus). Note that lesion 1 here is the same as lesion 2 in Figure 3.