(Mentor: Balaraman Ravindran)

The Correlated Topic Model (CTM) was proposed by Blei et. al to capture topic correlations. It does so by assuming the topics are drawn from a log-normal distribution. The log-normal and multinomial distributions are not conjugate, thus leading to complex inference and learning algorithms. We outline several approximations we explored to `collapse' the CTM to a simpler form that can be easily parallelised using a GPU. Parallelisation is key to scaling the model to large web-scale data.

• A working report of the various approaches we have attempted can be found here.