Sutra: Tensor-Op RNNs as a Compilation Target for Vector Symbolic Architectures
Researchers have developed Sutra, a functional programming language that compiles into PyTorch neural networks. This system targets vector symbolic architectures by reducing programs to fused tensor-operation graphs. Sutra demonstrates high accuracy in decoding bundles and allows for differentiable training directly through the compiled graph, enabling code to be both a logic program and a trainable neural network. AI
IMPACT Introduces a novel programming paradigm that unifies logic programming with neural network training.