A novel method for sampling a class of subsequential string transductions encoding homomorphisms allows rigorous testing of learning models' capacity for compositionality.
My doctoral dissertation, examining the relationship between abductive inference and algebraically structured hypothesis spaces, giving a general form for grammar learning over arbitrary linguistic structure.
We derive the well-studied subregular classes of formal languages, which computationally characterize natural language typology, purely from the perspective of algorithmic learning problems.
We comment on mathematical fallacies present in artificial grammar learning experiments and suggest how to integrate psycholinguistic and mathematical results.
We analyze the expressivity of a variety of recurrent encoder-decoder networks, showing they are limited to learning subsequential functions, and connecting RNNs with attention mechanisms to a class of deterministic 2-way transducers.