David Ha | Collective Intelligence for Deep Learning: A Survey of Recent Developments

Collective intelligence offers a transformative approach to deep learning, shifting from rigid, engineering-centric architectures toward adaptive, self-organizing systems. By drawing inspiration from biological and natural phenomena, researchers can develop models that respond dynamically to their environments rather than remaining indifferent to them. Neural Cellular Automata exemplify this by using local, stochastic rules to generate complex images and perform classification tasks with inherent robustness. Similarly, decomposing reinforcement learning agents into collections of communicating units enables zero-shot generalization across diverse morphologies and unseen challenges. Furthermore, modeling synapses as recurrent neural networks allows for the evolution of meta-learning rules that can surpass traditional backpropagation. These collective intelligence frameworks provide a promising path toward more resilient, flexible AI systems capable of learning to learn in complex, unpredictable settings.