M-Boost: Profiling and Refining Deep Neural Networks with Topological Data Analysis
Published in KDD 2018 Workshop on Interactive Data Exploration and Analytics, 2018
This extended abstract reports work in progress on a topology based approach to the problem of profiling, diagnosing and refining black-box models, with particular emphasis on deep neural networks. The proposed method is named M-Boost and relies on the mapper algorithm from topology, recursively identifying groups of observations where the accuracy can be improved.
Recommended citation: G Naitzat, N Lokare, J Silva, I Kaynar-Kabul, M-Boost: Profiling and Refining Deep Neural Networks with Topological Data Analysis. Workshop on Interactive Data Exploration and Analytics, KDD 2018, London, UK. http://poloclub.gatech.edu/idea2018/papers/idea18-paper10-naitzat.pdf