Residual Modeling for Time-Series Data using LSTMs (Jan 2017 - Present)
Technology: Python
Description: Working on a new algorithm for improved time-series modeling using a mixture of linear and non-linear models.

Malicious Domain detection from DNS server data (Nov 2016 - Present)
Technology: Python
Description: Working on an HMM/HsMM (Hidden (Semi) Markov Model) based framework for de-interleaving DNS server data and detecting new malicious domains using temporal information.

Multivariate Time-Series Modeling with Deep Learning (May 2016 - Oct 2016)
Technology: Python, Theano, Lasagne
Description: Time-series forecasting on aviation data. Performed careful empirical comparison between Vector Auto-Regressive (VAR) and Long Short-Term Memory (LSTM) based models (standard LSTM and sequential autoencoders) for time-series prediction. These are the plots of some sections of the original data and the predictions made by VARs and LSTMs. As can be seen from the plots above, VARs outperformed LSTMs in future time-series prediction. This can be attributed to multiple factors including the simplicity of the dataset and difficulty of training LSTMs. More details are discussed in the paper.

Head over to the Projects page to check out some other stuff I have done!