Dreams during rapid-eye-movement (REM) sleep may help learning and creativity through an adversarial process inspired by artificial intelligence, where a discriminator network classifies internally generated sensory activity as real. This process facilitates the emergence of real-world semantic representations in higher cortical areas and balances fantastic with realistic dream elements to promote creative insights. Non-REM (NREM) dreams, which replay single hippocampal memories, serve a complementary role by improving cortical representations' robustness to environmental perturbations. Subjects can become aware of adversarial REM dreams more easily than NREM dreams, and this awareness relates to wake, dream, and lucid dreaming phenomena.
Rapid-eye-movement (REM) dreams may function like an adversarial process—similar to generative adversarial networks in artificial intelligence—in which internally generated sensory activity is classified as real by a discriminator network. This adversarial dreaming helps build real-world semantic representations in higher cortical areas, supports learning, balances fantastic and realistic dream elements, and may spark creative insights. Non-REM (NREM) dreams, by contrast, replay single hippocampal memories, improving the robustness of cortical representations to environmental perturbations. The framework also explains differences in awareness of REM versus NREM dreams and how content- and state-awareness arise in wake, dream, and lucid dreaming states.