[Paper] Synaptic Activation and Dual Liquid Dynamics for Interpretable Bio-Inspired Models

Source: arXiv - 2602.13017v1

Overview

The paper introduces a unified, bio‑inspired framework for building and interpreting recurrent neural network (RNN) controllers. By extending classic RNNs with liquid‑capacitance dynamics and chemical‑synapse mechanisms, the authors achieve models that are not only high‑performing on a demanding lane‑keeping task but also far more transparent than typical black‑box policies.

Key Contributions

• Unified modeling language that captures a spectrum of bio‑inspired RNN variants (pure liquid dynamics, chemical synapses, and their combination).
• Liquid‑capacitance extension that endows dense all‑to‑all RNNs with interpretable internal dynamics without sacrificing expressiveness.
• Chemical synapse integration that further clarifies the causal relationship between network activity and control decisions.
• Comprehensive evaluation suite (turn‑weighted loss, neural activity‑trajectory correlation, saliency‑map robustness) applied to a realistic lane‑keeping control benchmark.
• Empirical evidence that the hybrid “chemical + activation” model delivers the best trade‑off between accuracy and interpretability.

Methodology

Model families

The authors start from a standard fully connected RNN and augment it in three ways:

• Liquid dynamics: each neuron behaves like a leaky integrator with a capacitance term, mimicking fluid‑like information flow.
• Chemical synapses: connections are modeled after neurotransmitter release, adding a non‑linear gating that can be turned on/off.
• Dual activation: both liquid and chemical mechanisms are combined, yielding a richer state‑space.

Training regime

All variants are trained end‑to‑end via reinforcement learning on a simulated autonomous‑driving environment that requires the vehicle to stay centered in its lane while handling curves of varying sharpness.

Interpretability metrics

• Turn‑weighted validation loss – penalizes errors more heavily on tight turns, highlighting where the controller matters most.
• Neural activity vs. trajectory correlation – measures how closely hidden states track the road curvature.
• Saliency maps – gradient‑based visualizations of which neurons influence the steering command at each timestep.
• Structural Similarity Index (SSIM) – quantifies the stability of saliency maps across similar driving scenarios.

Results & Findings

• The dual model reduces lane‑keeping error by ≈35 % relative to the vanilla RNN.
• Hidden‑state activity aligns much more closely with road curvature, indicating that the network is “thinking” about the same features a human driver would.
• Saliency maps become both clearer and more consistent (higher SSIM), making it feasible to audit the controller’s decisions post‑hoc.

Practical Implications

• Safer autonomous systems – Engineers can now deploy RNN‑based controllers with built‑in interpretability, facilitating regulatory approval and debugging.
• Debug‑first development – Saliency‑map stability lets developers pinpoint failure modes (e.g., loss of attention on sharp turns) before they manifest in the field.
• Transferable design patterns – The liquid‑capacitance and chemical‑synapse modules are lightweight plug‑ins that can be added to existing RNN libraries (PyTorch, TensorFlow) with minimal code changes.
• Explainable AI for control – The framework bridges the gap between high‑performance reinforcement‑learning policies and the need for human‑readable rationale, a key requirement in domains like robotics, aerospace, and medical devices.

Limitations & Future Work

• Scalability – Experiments are limited to a single lane‑keeping scenario; it remains to be seen how the approach scales to multi‑task or high‑dimensional perception pipelines.
• Computational overhead – Adding liquid and chemical dynamics modestly increases runtime and memory usage, which could be a bottleneck for real‑time embedded systems.
• Biological fidelity vs. engineering utility – While the models draw inspiration from neuroscience, the authors acknowledge that the abstractions are simplified and may not capture all relevant neuro‑dynamics.

Future research directions include extending the framework to vision‑based inputs, optimizing the bio‑inspired modules for edge hardware, and exploring formal verification techniques that leverage the newfound interpretability.

Authors

• Mónika Farsang
• Radu Grosu

Paper Information

• arXiv ID: 2602.13017v1
• Categories: cs.NE, cs.AI, cs.LG
• Published: February 13, 2026
• PDF: Download PDF