Resistance to Associative Interference in a Biologically-Constrained Active Inference Architecture: A Multi-Contingency Empirical Evaluation

This preprint presents an empirical multi-contingency interference test on the RDLNN architecture, directly addressing catastrophic forgetting — the fundamental failure mode of gradient-descent neural networks when learning sequential associations. Two Conditioned Stimuli (CS-A: visual; CS-B: auditory) are paired with independent Unconditioned Stimuli in strict sequential fashion across 130 continuous trials at 1 ms temporal resolution (585, 000 total ticks, 3 seeds). Primary Result: CS-A avoidance retention probe means of 0. 573, 0. 653, and 0. 523 across three seeds (pass criterion ≥ 0. 50: 3/3 seeds) after 30 trials of exclusive CS-B conditioning during which CS-A was never presented and never reinforced. The Lateral Amygdala weight matrix shows additive dual potentiation — a distinct rise at CS-A onset (Trial 30, Σ|W| rises to 23–24) and a second additive rise at CS-B onset (Trial 60, Σ|W| rises further to 38–40) — with no overwrite between them and stable retention through all probe trials. Three Structural Mechanisms of Non-Interference: (1) Orthogonal Liquid State Machine reservoir encoding — independent V1 and A1 reservoirs with separate random connectivity produce approximately orthogonal state trajectories for visual versus auditory inputs, generating orthogonal weight updates in the LA. (2) US-event-gated weight freezing — the isₜraining gate restricts LA updates to genuine shock events only; during CS-B acquisition CS-A weights are structurally frozen without any explicit protection mechanism. (3) 240× LTP: LTD asymmetry — even marginal cross-representation activation during the first CS-B trials cannot meaningfully depress CS-A weights within 30 trials at the safe-outcome LTD coefficient. Comparison with Standard Approaches: No replay buffers, elastic weight consolidation penalties, or architectural segregation are required. Non-interference emerges from the core learning rule structure rather than being imposed by external memory protection. The result is compared explicitly against EWC (Kirkpatrick et al. 2017), Progressive Networks (Rusu et al. 2016), and Experience Replay (Rolnick et al. 2019). Combined Continual Learning Case: Together with the preceding fear conditioning paper (extinction-not-erasure), this result closes the empirical continual learning argument: the architecture structurally separates both temporal interference (extinction) and associative interference (new contingency acquisition) without backpropagation, memory banks, or architectural tricks. Open Questions: Associative capacity limit (how many CS-US pairs before subspace overlap), simultaneous acquisition (both CSes presented in same trial), and matched US magnitude design to disambiguate magnitude effects from capacity effects. Keywords: catastrophic forgetting, continual learning, associative interference, fear conditioning, lateral amygdala, eligibility traces, active inference, variational free energy, liquid state machines, neuromorphic computing, complementary learning systems, biologically constrained learning Relation to prior work: Part of the RDLNN series. Integrated architecture paper: https: //doi. org/10. 5281/zenodo. 19186955. Fear conditioning paper: https: //doi. org/10. 5281/zenodo. 19217583.