Democratizing Learning with Neural Cellular Automata

Artificial Intelligence (AI) holds immense promise for transforming medical image analysis by providing faster and more accurate diagnosis, yet its benefits remain largely confined to wealthy regions. This disparity is particularly concerning in low- and middle-income countries (LMICs), where AI could reduce the burden on healthcare systems running at capacity but remains out of reach due to the high infrastructure demands of established models. This dissertation tackles the crucial challenge of democratizing AI by emphasizing efficiency, accessibility, trust, and adaptability, aiming to make advanced learning tools universally available. This dissertation focuses on bringing Neural Cellular Automata (NCA), a lightweight, efficient alternative to more conventional deep learning architectures, into the real world. Originally limited to toy problems, with the introduction of Med-NCA and M3D-NCA, NCA can now be applied to real-world tasks and handle high-resolution image segmentation. With 2-3 orders of magnitude fewer parameters than established architectures such as UNets, MED-NCA achieves similar accuracy, making them particularly suitable for resource-constrained environments. But, efficiency alone is not enough, accessibility is equally important. This research demonstrates that NCA can be deployed on minimal hardware, including Raspberry Pi devices and Android smartphones, bringing advanced AI capabilities into the hands of billions of people. The improved accessibility enables advanced healthcare solutions to extend beyond well-equipped facilities into underserved and remote regions. As models tend to fail on real-world data, ensuring trust is likewise a crucial challenge, especially in healthcare where reliability is critical. The NCA Quality Metric (NQM) is presented which provides quantitative feedback on model reliability, using the stochastic properties of NCA to signal potential failure scenarios. Finally, reducing dependence on large annotated datasets that are often unavailable in underserved regions is required to truly democratize learning. The Variance-Weighted Segmentation Loss (VWSL) approaches this problem, which is an adaptation method that enables the unsupervised finetuning of Med-NCA and adjusts to local data distribution shifts, removing the need for extensive labeled data. This dissertation aims to democratize learning by breaking down crucial barriers to efficiency, accessibility, trust, and unsupervised adaptability. By optimizing Neural Cellular Automata to handle high-resolution data, run on widely available hardware, detect failure cases, and finetune on unlabeled local data, this work aims to make access to advanced AI technologies equitable, bringing the benefits of AI into everybody's pockets, even in the most resource-constraint regions.