Despite remarkable advances in deep learning, contemporary AI systems remain brittle, data-hungry, and opaque. Concurrently, cognitive neuroscience has accumulated mechanistic insights (predictive coding, complementary learning systems, attention-based routing, and internal world models) that remain largely untapped by the AI community due to translation barriers. This paper provides a critical, conceptually grounded synthesis of these two fields in the form of a review and position statement. We identify foundational barriers to integration and discuss the most promising computational bridges, while acknowledging the historical precedent of biological metaphors that failed to scale. We argue for a Neurocognitive Framework (NCF) as a conceptual roadmap defined by four compatible design principles: (1) hierarchical predictive processing via local free-energy minimization, (2) fast and slow memory subsystems interacting through semantically-driven offline consolidation, (3) active top-down attentional gating modulated by prediction error, and (4) representational alignment with neural data as a regularizer and validation criterion. We review existing work instantiating these principles, critically examine the gaps that prevent their scalable implementation, and outline concrete, actionable steps for the community. We further propose NeuroCog-Bench, a dual-metric benchmark suite that jointly evaluates task performance and biological fidelity. We hypothesize, rather than predict, that significant advances in sample efficiency and robust generalization may require functional emulation of core cognitive motifs, and we document the conditions under which this hypothesis could be falsified.
Daniela A Pinzon Mora (Wed,) studied this question.