[R] Are neurons the wrong primitive for modeling decision systems?

A groundbreaking paper presented at the International Conference on Learning Representations (ICLR) challenges one of AI's foundational assumptions: that artificial neurons are the optimal primitive for building intelligent systems. The research team proposes 'Behavior Learning'—a novel architecture that replaces traditional neural network layers with learnable constrained optimization blocks. Instead of modeling intelligence through neuron activations and backpropagation, this approach frames decision-making as 'utility + constraints → optimal decision' processes. The authors argue that since many real-world systems (from economic markets to biological organisms) are fundamentally optimization-driven, AI architectures should reflect this reality at their core rather than approximating it through neural approximations.

The technical approach involves creating optimization modules that can be stacked and trained end-to-end, similar to neural layers but with fundamentally different internal mechanics. Each module takes inputs, defines a utility function and constraints, then solves for optimal outputs. Early experiments show these systems can learn decision policies with different characteristics than neural networks, particularly in tasks requiring explicit constraint satisfaction or multi-objective optimization. While critics might view this as 'structured inductive bias rebranded as a new paradigm,' proponents see it as a necessary evolution toward AI that better matches how real systems operate. The implications could be significant for fields like robotics, economics, and operations research where constraint satisfaction is paramount.