Program-as-Weights: A Programming Paradigm for Fuzzy Functions
The paper introduces "Program-as-Weights," a novel programming paradigm for defining and computing fuzzy functions. It proposes representing programs as weighted structures that map inputs to outputs with degrees of membership, offering a new approach to handling uncertainty in computation.
Background
- This paper introduces "Program-as-Weights," a programming paradigm where programs themselves serve as the weights (parameters) inside a neural network. Instead of learning numerical weights, the network learns which function to apply at each step.
- The approach targets "fuzzy functions" — functions that are neither purely symbolic (like traditional code) nor purely numeric (like standard neural nets), but a mix where the logic is partially unknown or approximate.
- This is a theoretical computer science / machine learning paper, not a deployed system. It sits in a line of work that tries to blur the line between programming and learning, such as neural-symbolic AI or differentiable programming.
- The key idea is that by making programs learnable parameters, you could get models that are more interpretable, compositional, and sample-efficient than standard deep learning — though the paper likely focuses on the formal definition and toy examples, not large-scale results yet.