Differentiable Exact Learning
Small neural controllers trained on tiny datasets yield discrete programs proven to perform binary arithmetic and universal Turing-machine simulation.
The framework pairs a state-tracking neural controller with an unbounded external environment and trains it on expert traces that reveal observations and actions but hide internal states. The authors report extreme length generalization with their Differentiable Finite-State Transducers, and reproduce that success with RNNs and GRUs. They extract interpretable discrete controllers from trained models and prove equivalence to the target algorithms. Exact-learning proofs for the neural controllers require assumptions that the authors support empirically.
ArXiv · AI/CL/LG's note
The framework pairs a state-tracking neural controller with an unbounded external environment and trains it on expert traces that reveal observations and actions but hide internal states. The authors report extreme length generalization with their Differentiable Finite-State Transducers, and reproduce that success with RNNs and GRUs. They extract interpretable discrete controllers from trained models and prove equivalence to the target algorithms. Exact-learning proofs for the neural controllers require assumptions that the authors support empirically.
ArXiv · AI/CL/LG's note
score 5