David Mareček presented his recent work on neural-networks interpretation, more specifically, what linguistic structures are latently learned by neural networks when solving downstream tasks like machine translation or sentiment analysis. Instead of using probing (supervised learning of linguistic features from hidden states), he focuses on an unsupervised extraction of the features. For instance, how part-of-speech tags are encoded in word embeddings or what syntax is learned inside the transformer’s self-attentions. See the whole presentation here: Searching for hidden linguistic structures in neural networks.