Safety in Reinforcement Learning
In Reinforcement Learning the agent explores the environment and changes its behavior to adapt to the environment dynamics. In such a situation, where the agent is modifying its behavior, safety becomes a critical component in order to avoid dangerous or inappropriate behavior. In this work, used a directed exploration technique which incorporated a notion of safety along with maximizing the cumulative rewards.
Participated in ICLR 2018 Reproducibility Challenge. The aim was to ensure the accuracy of the claims and results of the paper.
CNN for arithmetic character recognition
Using Deep Net methods for mathematical character recognition in image
The aim of this project is to use various machine learning algorithms to classify an image with two handwritten digits, and one handwritten character with their arithmetic results. With a convolutional
neural network of 4 layers of convolution blocks, a score of 92:8% is achieved on the Kaggle competition.
Anomaluy Detection in real- time Cloud Services
I developed a service diagnostic and anomaly detection tool. Traditional heuristics based data analytics approaches fail in joint modelling of multivariate temporal system performance signal. We devised a generic framework for unsupervised hierarchical diagnosis of services. There was a dearth of labelled data, therefore developed a Bayesian model for cloud service that created synthetic time-series emulating characteristics
of high-dimensional real data. Individual time series were segmented and labelled to identify different levels of issue severity based on anomalous behavior. Then, I created objective function utilizing Viterbi algorithm for approaching segmentation and labelling in a joint model to group the key anomalous patterns across the service.
Language Classification in Mixed Data via character recognition
Classifying sentence based on individual character. In order
to do this, features were computed at character level rather than word level.