Mettre l’intelligence à l’échelle : un cadre formel et pratique pour une IA à capacité de calcul non bornée

Fixed AI architectures face computational limits, as complexity theory shows, with problemsrequiring exponentially scaling resources. This dissertation presents Computationally Unbounded AI(CUAI), enabling autonomous capacity scaling inspired by neural modularity and plasticity. CUAIovercomes static model fragility in dynamic secngs via self-directed expansion. Key contribu)onsinclude: (1) the Column Extension Framework, which scales ANNs through graph rewriting,duplicating subgraphs for growth, validated on MNIST-1D; (2) the Capacity-Aware Learning (CAL)Framework, combining boosting and selective prediction with MLP auto-encoders to enhance taskcoverage and accuracy, outperforming single models. Drawing from cortical columns, CUAI fostersself-improving systems. Future work targets real-world applications and alignment for autonomous,scalable intelligence.