A vast variety of sensory cues are represented as distinct patterns of neural responses in the brain. To retrieve memories associated with sensory cues, one would assume that the representations of sensory signals should be stable over time. However, recent development of population-level recording techniques has revealed that these representations can drift over time $^{1}$. To investigate the mechanisms and consequences of representational drift, we formulate a random feedforward network model $^{2,3}$ that incorporates stochastic fluctuations in membrane potential and synaptic strength $^{4,5,6}$. Our findings show that while both types of fluctuations similarly induce drifts in individual odor representations, they have different effects on the discriminability between representations; fluctuations in potential reduce, while those in synaptic strength preserve it. We explain the underlying mechanisms of discriminability preservation by using the self-averaging property inherent in random networks. Using a reversal learning task, we demonstrate that both types of drift can enhance learning efficiency for associating a single odor-valence pair, provided that the drift rate is appropriately matched to the reversal rate. However, only drift induced by synaptic fluctuations improves associative learning of multiple odor-valence pairs, as it requires discrimination between odor representations. We derive the optimal drift properties that maximize the signal-to-noise ratio using the Ito isometry and the Cauchy-Schwarz inequality, providing a theoretical basis for the observed improvements in learning. Our findings shed light on the source of representational drift and its potential benefits for flexibility of learning under dynamic environments.