Why human vision needs both generative and discriminative learning?

A persistent debate asks whether human visual perception maps onto discriminative or generative learning. Using Joint Energy-Based Models to smoothly interpolate between these objectives within a single architecture, researchers tested six human-alignment benchmarks—perceptual similarity, robustness, shape-texture trade-offs, and others. The key finding: alignment peaks at intermediate mixing ratios, not at pure endpoints. Hybrid models capture both the categorical structure of discriminative learning and the input sensitivity of generative learning, suggesting the true recipe for human-like vision balances both.