Emergent Fear Conditioning Dynamics in a Biologically-Constrained Ten-Module Active Inference Architecture

Read full text Summarise with VIRAT
Authors: Anol Deb Sharma
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This preprint demonstrates emergent, biologically-plausible fear conditioning behaviour in a continuous ten-module active inference architecture (RDLNN) without global backpropagation or explicit programming of any behavioural target. A 131-trial four-phase paradigm — Habituation, Acquisition, Extinction, and Reinstatement — is run continuously at 1 ms temporal resolution across three independent seeds (589,500 total ticks). Primary Result — Extinction-Not-Erasure: The Lateral Amygdala weight tensor shows zero magnitude through 50 habituation trials, sharp potentiation to Σ|W| ≈ 23–24 at acquisition onset, a stable plateau maintained through all 30 extinction trials with zero decay, and further potentiation to 34–36 at the reinstatement probe. The ITC pathway (PFC-driven inhibitory gating) suppresses avoidance behaviour during extinction without modifying the underlying LA associative weights. This is the direct computational implementation of the biological finding that extinction is new learning rather than unlearning. Additional Confirmed Signals: Extinction behavioural suppression (AvoidRate drops to 0.21–0.58 at Trial 110; VFE drops from ≈64 to 35–51 at phase boundary), partial reinstatement recovery (AvoidRate rebounds to 0.49–0.63 after unsignalled US probe), and acquisition S-curves confirmed in two of three seeds. Seed 100 acquisition variance is documented as BG initialisation sensitivity rather than architectural failure. Five Architectural Invariants: Five invariants that only manifest under continuous 1 ms chronological execution are documented: ITI noise erasure (C1), eligibility trace accumulation collapse (C2), 1 ms motor entanglement (C3), LTP/LTD rate symmetry producing the Avoidance Paradox (C4), and f-expected integrator timescale saturation (C5). None are derivable from the architectural specification. All emerged under continuous multi-module execution. Open Limitation: VFE remains near-saturated within the 30-trial acquisition window at current f-expected τ = 50 ms. The resolution path (hierarchical VFE decomposition separating CS and US surprise channels) is specified for the Real_build phase. Relation to prior work: Part of the RDLNN series. Integrated architecture paper: https://doi.org/10.5281/zenodo.19186955

Fields of Study & Subjects

Memory and Neural Mechanisms Embodied and Extended Cognition Psychiatry Mental Health Neuroscience

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint