This is the repository for my final project for course CS242: Probabilistic Graphical Models (CS department at Brown University).

http://cs.brown.edu/courses/cs242/lectures/

Title: Bidirectional vision through Deep Boltzmann Machine boosted with conjugate neural networks

Abstract:

We created a hierarchical MRF (deep Boltzmann machine, or DBM) that does both discriminative classification (bottom-up) and generative sampling (top-down) tasks, as the bidirectional biological visual system. Compared with the original DBM, we used two novel techniques to improve the performance: 1) the importance weightening effectively balanced the layers and help to implement a sufficient amount of top-down influence; 2) We boosted the DBM learning using the response of a conjugate feed-forward neural network, the improved efficiency makes it possible to be applied to very deep networks.

Full text report, see file final_report in the root directory.