Resilient functioning in the face of sensory failure is a characteristic of biological systems, but remains a challenge for artificial neural networks. We subjected a compact recurrent network (262k parameters) to one hour of total sensory deprivation coupled with a block on plasticity, followed by the restoration of both inputs and plasticity. Using fractal metrics (Hurst exponent H and fractal dimension D), global uncertainty, and a consistency index C, we discovered two phenomenological laws governing post-deprivation recovery. AUNI Law Law of SCO AUNI's law describes a rapid exponential restoration of the internal fractal structure (τ ≈ 5.3 min for D, and τ ≈ 5.8 min for uncertainty).On the other hand, SCO's law reveals a strong inertia in the predictive consistency of system C, which remains stable during deprivation and recovers only very slowly (τ > 12 min).This dual dynamic—rapid microstructural repair versus slow macrocognitive adaptation—demonstrates a functional dissociation within the same artificial cognitive system. Our results provide the first quantitative experimental validation of such a dissociation in a recurrent neural network, opening new avenues for the design of resilient and self-repairing AI. This represents a complete paradigm shift.
Remi SCOGNAMIGLIO (Tue,) studied this question.