Between Amnesia and Chaos: A Memory Stability Expressivity Trilemma for Trainable Dissipative Oscillator Networks
Researchers have identified a fundamental trilemma in training physical reservoir computing systems, specifically dissipative oscillator networks. The study reveals that maximizing memory horizon, gradient stability, and dynamical expressivity simultaneously is impossible, as all three are dictated by the network's damping. This damping level critically affects how far back in time gradients can propagate and influences the stability of forward sensitivities, confining usable training to a specific band that narrows with increasing memory requirements. AI
IMPACT Identifies a fundamental constraint in training physical AI systems, potentially guiding future research in neuromorphic computing and hardware-based AI.