Hari Shrawgi is an MTech AI student at the Indian Institute of Science where he is working to improve the theoretical foundations of deep learning. He got interested in ML during his under-graduation, doing extensive applied research which led to five research publications. Working on sensitive applied domains such as biomedicine and gene-editing, he realized the need for trustworthy and explainable AI. This need underpins his research goal - to develop theory that can lead to more trustworthy AI models. In his free time, he loves saving virtual worlds in video games.
Interned with the ‘Trustworthy Fundamentals’ team at Bing which focuses on issues around Trust and Fairness. Majority of my time as an intern was devoted to building models to detect search queries that can spread political misinformation and hatred. I also developed a twitter bot using these models to detect tweets which may lead to misinformation.
Worked as an R&D Engineer for the product CA Single Sign-on under the security division. It is a product that serves most of the Fortune 500 companies and is deployed on the biggest scales of enterprise software in the world. Following are the highlights of my contribution to the product and the company:
Worked under the guidance of Prof. Brett Lidbury in the interdiscilinary field of Bioinformatics:
Developed a new technique to reduce training sample requirements for Hierarchical RL algorithms. The technique recycles the data collected at a lower temporal scale for training higher temporal layers. The technique is very easy to plug into most existing and widely used HRL algorithms. The techinque was observed to reduce the training data requirements for HDQN by 50% in a simple MDP setup.
The CRISPR/Cas9 system for gene editing relies heavily on the selection of a good RNA guide. The manual selection process is both difficult and expensive. As part of my B. Tech major project, I developed a CNN model to automate this process. Below are short pointers related to the project: The model was trained on over 400,000 data points from approx. 400 human cell lines. The model outperformed all conventional machine learning models which are dependent on feature enineering.
View Publication related to this project.Used Neural Networks coupled with ROC curve analysis to discover biomarkers for clinical diagnosis of CFS. Substantiated the results with a systematic review on 60 NCBI research works and articles. The new biomarkers can lead to detection of CFS through pathological tests which was not feasible before.