Synaptic and potential fluctuations drive representational drift with differential effects on discrimination learning

Large-scale neuronal measurements have revealed continual drift of neural representations without explicit learning, raising fundamental questions about how the brain maintains reliable perception. While representational drift has been studied both experimentally and theoretically, the underlying mechanisms and computational consequences remain unclear. Here, we systematically compare two drift mechanisms--fluctuations in synaptic weights versus membrane potential--using an analytically tractable neural network model with aligned drift magnitude and timescale. We found that both mechanisms preserve representational structure across time regardless of the drift magnitude. However, only synaptic fluctuations maintain discriminability between representations under large drift magnitude. We use a selectivity space to explain how these similar and differential effects on representational drift emerge from the dynamics of individual neuronal tuning. We further examine functional consequences of drift using reversal learning tasks where stimulus-reward associations switch. Both mechanisms enhance single stimulus learning by gradually removing the old associations, but only synaptic fluctuations substantially improved learning of multiple stimuli by preserving the representational structure during drift. Our results reveal distinct computations performed by specific fluctuation mechanisms and explain how representational drift can support both stable and adaptive behavior.