Modeling focus state and mind-wandering with predictive-coding-inspired VRNN

• Autonomous shifts between focus state and mind-wandering have been modeled using a predictive-coding-inspired variational recurrent neural network (VRNN).¹
• The computational model could inform neurofeedback or state-dependent BCI; it does not use direct electrophysiology.¹
• Frontiers-published; tier-2 watchlist for computational neuroscience.^{1 1}

Weekly enrichment (2026-07-20)

• The study extends the Predictive Coding-inspired Variational RNN (PV-RNN) of Ahmadi and Tani (2019) by adding an online adaptation mechanism for the meta-prior w, the parameter that weights the complexity term against the accuracy term in free-energy minimization.^{2 3}
• It was published in Frontiers in Computational Neuroscience on 2 July 2025 (Oyama, Matsumoto & Tani, Okinawa Institute of Science and Technology) and is a pure model-simulation study with no human electrophysiology or participant sample.^{2 3}
• In simulations a low w emphasized bottom-up sensory input and corresponded to the focus state, while a high w prioritized top-down internal predictions and corresponded to mind-wandering.^{2 3}
• Autonomous shifts emerged when w switched between low and high values in response to the average reconstruction (prediction) error over a fixed-size past window—high w when error fell near a minimal threshold, low w when it rose near a maximal threshold.^{3 4}
• The authors speculate that conscious self-awareness of mind-wandering occurs only once accumulated error exceeds a threshold, matching the empirical asymmetry that focus-to-wandering shifts happen unconsciously while wandering-to-focus returns require conscious recognition.^{2 3}
• The model builds on Friston’s free energy principle and earlier PV-RNN allostasis work (Idei et al., 2024); unlike those studies, transitions here are generated autonomously rather than by manually resetting w.³
• No numeric accuracy benchmarks against human data are reported, because the paper evaluates simulated cyclic sensory sequences rather than recorded neural signals (thin on empirical metrics).^{2 3}
• For BCI context, EEG-based mind-wandering decoders in humans reach roughly 80-83% within-subject accuracy with common spatial patterns, and AUC ~0.88 within-participant versus ~0.70 cross-lecture with Riemannian-geometry SVMs, illustrating the gap a generative state model could help close for state-dependent neurofeedback.^{5 6}

Footnotes

1. https://news.google.com/rss/articles/CBMipwFBVV95cUxNZUF1aFhmVHR2bWJxMFBIb1Z2emJUUklUQVNCX3pZZWVGTUZHYmx5alo4RUFJY2hFMDJEQ1d0bUhDbTRUVzl1LV9HOXlMRUU5cXkxdGVSSHQ4Z3BDOFdtTDJISjJmbGdqUDBkV2pGdElRU3YtNWxsOXdmcmxjaS1ZQzhDZjFhQUVTRUdyWkpTRmx1OXRhNHhkSFlYT0dsOVBHTUJjUDVNYw?oc=5 ↩ ↩² ↩³ ↩⁴

2. https://www.frontiersin.org/journals/computational-neuroscience/articles/10.3389/fncom.2025.1578135/full ↩ ↩² ↩³ ↩⁴ ↩⁵

3. https://pmc.ncbi.nlm.nih.gov/articles/PMC12263957/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

4. https://arxiv.org/abs/2412.15620 ↩

5. https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0222276 ↩

6. https://pmc.ncbi.nlm.nih.gov/articles/PMC10267732/ ↩