English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Tools annual

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

Natural languages are powerful tools wielded by human beings to communicateinformation. Among their desirable properties, compositionality has been themain focus in the context of referential games and variants, as it promises toenable greater systematicity to the agents which would wield it. The concept ofdisentanglement has been shown to be of paramount importance to learnedrepresentations that generalise well in deep learning, and is thought to be anecessary condition to enable systematicity. Thus, this paper investigates howdo compositionality at the level of the emerging languages, disentanglement atthe level of the learned representations, and systematicity relate to eachother in the context of visual referential games. Firstly, we find that visualreferential games that are based on the Obverter architecture outperformsstate-of-the-art unsupervised learning approach in terms of many majordisentanglement metrics. Secondly, we expand the previously proposed PositionalDisentanglement (PosDis) metric for compositionality to (re-)incorporate someconcerns pertaining to informativeness and completeness features found in theMutual Information Gap (MIG) disentanglement metric it stems from. Thisextension allows for further discrimination between the different kind ofcompositional languages that emerge in the context of Obverter-basedreferential games, in a way that neither the referential game accuracy norprevious metrics were able to capture. Finally we investigate whether theresulting (emergent) systematicity, as measured by zero-shot compositionallearning tests, correlates with any of the disentanglement and compositionalitymetrics proposed so far. Throughout the training process, staticallysignificant correlation coefficients can be found both positive and negativedepending on the moment of the measure.

Visual Referential Games Further the Emergence of Disentangled Representations