Modeling Autonomous Shifts Between Focus State and Mind-Wandering Using a Predictive-Coding-Inspired Variational RNN Model

TL;DR

The paper addresses how autonomous shifts between focused attention and mind-wandering can arise from predictive-coding dynamics. It extends a Predictive Coding Inspired Variational RNN (PV-RNN) by introducing a meta-prior $\mathbf{w}$ that adapts based on past reconstruction error to bias inference toward bottom-up or top-down processing, enabling FS and MW without manual switching. Through training on cyclic sensory patterns and testing with dynamic meta-prior switching, the authors demonstrate that low $\mathbf{w}$ yields accurate reconstruction (FS), while high $\mathbf{w}$ promotes top-down imagery and wandering (MW), with transitions governed by a temperature parameter. The work links free-energy minimization with autonomous attentional shifts and outlines paths to connect the model with resting-state networks and conscious awareness mechanisms.

Abstract

The current study investigates possible neural mechanisms underling autonomous shifts between focus state and mind-wandering by conducting model simulation experiments. On this purpose, we modeled perception processes of continuous sensory sequences using our previous proposed variational RNN model which was developed based on the free energy principle. The current study extended this model by introducing an adaptation mechanism of a meta-level parameter, referred to as the meta-prior $\mathbf{w}$, which regulates the complexity term in the free energy. Our simulation experiments demonstrated that autonomous shifts between focused perception and mind-wandering take place when $\mathbf{w}$ switches between low and high values associated with decrease and increase of the average reconstruction error over the past window. In particular, high $\mathbf{w}$ prioritized top-down predictions while low $\mathbf{w}$ emphasized bottom-up sensations. This paper explores how our experiment results align with existing studies and highlights their potential for future research.

Figures (7)

• Figure 1: A hierarchical two-layer PV-RNN architecture. Solid blue lines represent the generative process, while dotted red lines indicate the inference process. The shaded area shows an inference window of length 3.
• Figure 2: Training trajectory over 400 time steps (top plot) and its representation in $X-Y$ space (bottom plot). $Target - X$ and $Target - Y$ correspond to the first and second dimensions of the training trajectory, respectively.
• Figure 3: Prior generation over 1200 time steps under trained model with meta-prior $\mathbf{w}^{tr}$ from Table \ref{['tab:parameters']} (top plot), selected activities of the $\mathbf{d}$ neurons in the bottom layer of the PV-RNN (middle plot), and a representation in $X-Y$ space (bottom plot). $Output - X$ and $Output - Y$ correspond to the first and second dimensions of the prior generation output trajectory, respectively.
• Figure 4: From top to bottom: inference output trajectory with meta-prior $\mathbf{w}^{L}$ from Table \ref{['tab:param_test']}, selected activities of the $\mathbf{d}$ neurons in the bottom layer of the PV-RNN, average reconstructions error over the inference window at time step 542, and KL divergence at the PV-RNN bottom layer. $Inference - X$ and $Inference - Y$ correspond to the first and second dimensions of the inference output trajectory, respectively.
• Figure 5: From top to bottom: inference output trajectory with meta-prior $\mathbf{w}^{H}$ from Table \ref{['tab:param_test']}, selected activities of the $\mathbf{d}$ neurons in the bottom layer of the PV-RNN, average reconstructions error over the inference window at time step 283, and KL divergence at the PV-RNN bottom layer. $Inference - X$ and $Inference - Y$ correspond to the first and second dimensions of the inference output trajectory, respectively.