How to grow a self: development of self-representation in the Bayesian brain

Abstract

The last two decades saw multiple attempts to explain how the self is represented in the brain within the framework of the Bayesian brain. However, these attempts largely focused on describing a developed, adult self-representation. The current paper argues that looking at the developmental trajectory is crucial for understanding the structure of self-representation. It argues that emergence of self-representations should be understood as an instance of the process of acquisition of new internal models of hidden causes of sensory input. The paper proposes how such models emerge and develop over the course of life by looking at different stages of development of bodily and extra-bodily self-representations. It argues that the self arises gradually in a series of discrete steps: from first-person multisensory representations of one’s body to third-person multisensory body representation, and from basic forms of the extended and social selves to progressively more complex forms of abstract self-representation. It discusses how each of them might emerge based on domain-general learning mechanisms, while also taking into account the potential role of innate representations. Finally, it discusses how predictions of the proposed model might be experimentally tested.

Keywords: Bayesian brain; abstract self; bodily self; cognitive development; predictive coding; self; self-recognition; self-representation.

Figure 1

An example of a localist connectionist model of self-representation based on the associative network model of memory (cf. Kihlstrom et al., 2003; Smith et al., 1999). Mental representations are presented as nodes on a two dimensional plane spanning perceptual-to-abstract content and semantic-to-episodic content. The internal model of the self and a friend are visualized as elements of the semantic memory. These nodes are connected with other representations. Thicker lines denote strong connections between the nodes, while thinner lines represent weaker ones.

Figure 2

Self-representation as a hierarchical structure of internal models with the self-model as a high level abstract prior (on the left). Each link in the hierarchical structure reflects the interplay of top-down and bottom-up signals as illustrated by the link in pink. In Bayesian terms the top-down signal reflects prediction or strength of the prior, and the bottom-up direction reflects prediction error or the incoming data. The self-model (the blue node) is connected to the representation of a person’s voice and face, but also to the representation of one’s eye illustrating that representations are free to be connected across levels of the hierarchy.

Figure 3

A graphical illustration of the process of acquisition of the concept of a natural number (NN). According to Carey (2009) children are born with an innate concept of magnitude (through “Analog Magnitude System,” or AMS) which allows them to differentiate between, e.g., long and short objects. Children are also born with a “Parallel Individuation System” through which during the third year of life children begin to understand the concept of individual numbers: one, two, three, sometimes four. The crucial point of Carey’s proposal is that at certain point children discover that individual numbers can be mapped onto the underlying representation of magnitude (Carey proposes that it happens through the mechanism of Quinian bootstrapping) and hence acquire the concept of a natural number.

Figure 4

Graphical illustration of acquisition of four types of self-models. In each case the existing representations/models (shown on the left in each section) are recombined (indicated by the yellow arrows) into novel self-models (shown on the right). (A) The primordial self-model emerges once the brain discovers that it can classify sensory signals into the ones which are reliably predicted by one’s motor behaviour and those which are not. (B) The same classification rule can be applied to visual perception leading to acquisition of first-person representation of one’s body. (C) A toddler that possesses the concept of another person and understands that some visual percepts represent their body can discover that in special cases (e.g., encounter with a mirror) they can come together, and that it is possible to see oneself from the third-person perspective. (D) An abstract self-model emerges once a child discovers that self/non-self distinction can be applied to abstract representations.

Figure 5

A schematic illustration of emergence of the primordial self-model. (A) At the outset an organism has only an implicit model of sensory noise. (B) The discovery that some sensory stimulation reliably co-occurs with one’s motor commands leads to the emergence of the primordial models of self and else (see also Figure 4A).

Figure 6

Theoretically possible trajectories of change of self-concept. However, as described in the text, empirical evidence favors a variant of model B2.

Figure 7

Overview of proposed stages of development of self-representations.