Sagar Garg

Hi I’m Sagar. Here are some of my projects

AI + Weather (Computer Vision)

WeatherBench Probability was my masters thesis work, built on extending the existing WeatherBench project by European Centre for Medium range Weather Forecasting (ECMWF). I extended the project from deterministic to probabilistic approaches. The objective is to find weather predictions 3,5 days ahead. Since Weather is a chaotic phenomena, we proposed 3 ensemble methods to take care of the inherent uncertainty in predictions. Paper published. Poster at EGU Conference/ One-Slider.

AI + Hydrology (Time-Series)

We won 2nd prize at the 6-month Wave2Web hackathon by World Resources Institute in association with Microsoft to build Machine Learning Solutions for Reservoir modelling. Our product aims to help governments forecast water availability in reservoirs 30,60,90 days ahead, making it a crucial knowledge asset to cope with water scarcity. Poster at AGU Conference/ One-Slider

AI Safety

I participated in the Artificial General Intelligence Safety Fundamentals Programme by Effective Altruism Cambridge, with Technical Alignment curriculum designed by Richard Ngo, DeepMind. Code Projects that I have done in this domain include: Robustness, Adversarial attacks and Certification, Uncertainty-aware Probabilistic Prediction. This is a topic of extreme seriousness for the future, and I am happy to learn and talk about it more.

Generative AI

Research on Generative Adversarial Networks to improve robust performance of predictive models suffering from small data. The exciting part was to answer the question - How do you find good generators, which lie in a tiny subspace of a high-dimensional space?

AI Model Automated Optimization and Re-Training

Worked for an enterprise-critical project to build the SOP for the training of all machine learning models, specifically with regards to hyperparameter tuning, and model periodic re-training.

Carbon in Indian Forests

I Analyzed remote-sensed satellite image data (radar and optical) using GIS tools to estimate carbon in Indian forests over 30 years. This was my time at Divecha Centre for Climate Change, Indian Institute of Sciences, Bengaluru. I attended all lectures, seminars and courses possible to learn as much about the field as possible. Deciding to do more, I chose my masters degree. Interesting fact: An isolated forest area is much better at sequestering carbon than the same area with disturbance such as roadlines nearby.