AI Learns and Self-Assesses Using 'Surprise' Signal

Researchers are investigating a novel concept: using a "surprise" signal, derived from prediction errors within a frozen encoder's latent space, to drive both learning plasticity and metacognition in AI. This single idea was explored across two distinct systems.In the first system, a non-parametric episodic memory only records new concepts when this surprise signal is high. A periodic offline replay phase then consolidates these recent memories into a long-term store. This approach significantly improved retention on older classes in a continual learning setting with ImageNet, recovering substantial performance points compared to baselines.The second system applied the same surprise signal, but in a shared text-image space, to modulate a vision-language model's behavior. This model could confidently answer when a concept was known, express uncertainty (hedge) when partially familiar, or refuse to identify an object and request an explanation when it encountered something truly novel. This external surprise detector proved far more accurate than the model's own verbalized confidence in distinguishing known from novel concepts. After a "sleep" phase, the system recalled nearly all taught facts from its consolidated store, demonstrating robust long-term memory.