Evolving Afferent Architectures: Biologically-inspired Models for Damage-Avoidance Learning

TL;DR

This paper tackles long-horizon damage-avoidance learning by introducing Computational Afferent Traces (CATs), internal risk signals produced by evolved afferent arrays. It presents a bi-level architecture where outer-loop evolution via CMA-ES discovers afferent sensing configurations and inner-loop PPO trains damage-avoidance policies using CATs, formalizing an inductive bias toward learnability. The approach is demonstrated on biomechanical digital twins, showing evolved CAT architectures achieve 2.8x CAT efficiency and 15.4x age-robustness versus hand-designed baselines, along with a 23% reduction in high-risk actions; ablations confirm the necessity of CAT signals, evolution, and predictive discrepancy. Episodic memory (AMM) further enhances adaptation, and generalization across pathological knee conditions with age-dependent action restriction is demonstrated. The work provides a reproducible framework and open data to advance internal risk signaling in AI for biomedical and embodied systems.

Abstract

We introduce Afferent Learning, a framework that produces Computational Afferent Traces (CATs) as adaptive, internal risk signals for damage-avoidance learning. Inspired by biological systems, the framework uses a two-level architecture: evolutionary optimization (outer loop) discovers afferent sensing architectures that enable effective policy learning, while reinforcement learning (inner loop) trains damage-avoidance policies using these signals. This formalizes afferent sensing as providing an inductive bias for efficient learning: architectures are selected based on their ability to enable effective learning (rather than directly minimizing damage). We provide theoretical convergence guarantees under smoothness and bounded-noise assumptions. We illustrate the general approach in the challenging context of biomechanical digital twins operating over long time horizons (multiple decades of the life-course). Here, we find that CAT-based evolved architectures achieve significantly higher efficiency and better age-robustness than hand-designed baselines, enabling policies that exhibit age-dependent behavioral adaptation (23% reduction in high-risk actions). Ablation studies validate CAT signals, evolution, and predictive discrepancy as essential. We release code and data for reproducibility.

Figures (6)

• Figure 1: Knee afferent arrays. Labels A1--A6 indicate six functionally specialized arrays (see Figure \ref{['fig:array_specialization']} and Table \ref{['tab:array_anatomy']}).
• Figure 2: Brick layer lifetime progression.
• Figure 3: Comparison of evolved vs. baseline afference configurations ($n=5$ runs).
• Figure 4: Afferent array specialization: average feature weights per array. Red = positive (activating), blue = negative (inhibitory).
• Figure 5: Ablation studies ($n=10$ runs per variant). Variants: no_CAT (direct damage feedback), no_evolution (hand-designed), no_amm (episodic memory disabled), no_predictive (CAT without predictive discrepancy).

Theorems & Definitions (1)

• Definition 1: Computational Afferent Trace (CAT)