How do neural language models keep track of number agreement between subject and verb? We show that `diagnostic classifiers', trained to predict number from the internal states of a language model, provide a detailed understanding of how, when, and where this information is represented. Moreover, they give us insight into when and where number information is corrupted in cases where the language model ends up making agreement errors. To demonstrate the causal role played by the representations we find, we then use agreement information to influence the course of the LSTM during the processing of difficult sentences. Results from such an intervention reveal a large increase in the language model's accuracy. Together, these results show that diagnostic classifiers give us an unrivalled detailed look into the representation of linguistic information in neural models, and demonstrate that this knowledge can be used to improve their performance.As always, after the talk there will be beers and snack in the common room.

We introduce the graphical language of string diagrams, which allow us to reason about mathematical structures by drawing pictures. Specifically, we use string diagrams to define monoids and comonoids, and demonstrate how the matrices of natural numbers arise from interactions between a monoid and a comonoid. By studying a certain class of categories known as PROPs, we will see that the diagrammatic approach is in fact in one-to-one correspondence with the algebraic one. We proceed to outline how this generalises to matrices with rational entries, thus recasting all of the (rational, finite-dimensional) linear algebra in terms of string diagrams.

Join us for snacks and drinks in the common room after the talk!

The ILLC has a new group, called the Causal Inference Lab, dedicated to investigating causality. The group hosts a reading group every second Monday, 13:00-15:00, which everyone very welcome to attend.

For our next session, Monday 18 February 13:00-15:00, we will read Rachael Briggs' paper, 'Interventionist Counterfactuals' (https://doi.org/10.1007/s11098-012-9908-5). Briggs applies causal models to represent the truth-conditionals of counterfactuals such as "If you had taken the train or the metro, you would have arrived on time."

If you are interested in discussing causal inference with us, please do come along!

We interact with computers every day, and often using something like human language. There is therefore a huge amount of research going into how to represent human language computationally. Modelling words as vectors has been one of the most successful approaches over recent years. However, it is not immediately clear how to combine word vectors together to make phrases and sentences. On the other hand, formal semantics gives a clear
account of how to compose words, but it is not so obvious how to represent their meanings. I will give an overview of the model I work with that shows how to combine word vectors using formal semantics. I will also describe its limitations and will appreciate ideas and questions.

Abstract:
In contemporary Western philosophy, destruction or deconstruction of metaphysics has been presented by prominent thinkers like Levinas, Derrida or Schürmann as the liberation of an an-archic way of thinking. The possibility of questioning and acting beyond the «arkhè», beyond the principle (commandment and beginning at the same time), has opened new perspectives in ontology and ethics. Levinas, for example, characterizes the relation to the Other in terms of «an-archic responsibility». Interestingly, such an ontological and ethical anarchy has always been strictly distinguished from political anarchism. Can we envisage a confrontation between the two traditions? Is "post anarchism" a satisfactory way of setting it up?