Hi, I’m Michael Goodale. I’m a PhD student working at Institut Jean-Nicod at École normale supérieure where I am advised by Salvador Mascarenhas and Yair Lakretz. My thesis looks to apply different inference algorithms to the formalisms that have been developed in theoretical linguistics in order to model language acquisition as well as ways of connecting formal linguistic representations with what is learnt by neural network models. This involves learning a Montague grammar from small data where the propositions paired with sentences are strictly latent.

I am interested in how human thought and neural networks’ abilities diverge, and how they might be reconciled. In particular, I have done research into how neural networks are and aren’t systematic (in the sense of Fodor and Pylyshyn). Systematicity refers to the human property of being able to perform cognitive tasks without being perturbed by irrelevant noise and our ability to combine different abilities together even if we’ve never seen them combined. For example, any human who understands the sentence “dog bites man” must also understand “man bites dog”, despite its statistical improbability.

Some of my research looks to see whether neural networks are systematic and investigates different approaches to make them systematic. I have particularly looked at Meta-Learning approaches (see Goodale, Mascarenhas and Lakretz 2025 or Goodale and Mascarenhas, 2026), but I personally think neuro-symbolic approaches are most promising.

I did a master’s degree in cognitive science at ENS where I developed a model of conceptual processing where the core primitives of formal semantics, discrete sets, are replaced with continuous representations, namely manifolds. This was to develop a novel account of privative adjectives, generic statements, and classical reasoning errors like conjunction fallacy. You can find a copy of it here on lingbuzz.

You can download my full résumé here.