Projects

Description

Investigating CoT, self-play reasoning and coding capabilities in RL-trained LLMs. This is a Master Thesis Project at the University of Tübingen.

Description

Investigating properties and reward functions of GRPO for training small LLMs on a toy math dataset. This is a Research Project at the University of Tübingen.

Description

This is a course project at the University of Tübingen.

Abstract: Weather-dependent energy output forecasting systems have suffered due to the inherent noise in weather elements. Diffusion models can improve such forecasting by learning these noise patterns. In this work, we show that diffusion models can be used on top of any regression-based model for better forecasting. The quality of forecasting is improved by the denoising diffusion-based conditional generative model over a pre-trained conditional mean estimator. The conditioning prior model and the diffusion model are trained and tested using the RMSE metric on a dataset containing 4 years of hourly weather data with corresponding solar and wind energy supply values. It is observed that the diffusion mechanism consistently outperforms the initial conditioning prior by 4-5% on average.

Description

Using multimodal data with graph neural networks for small molecules and SOTA LLM-based protein and molecule encoders for predicting mechanism of action (MoA) in drug-target interactions. The prediction of MoA is a novel work, never done before. We also use secondary structure information of protein as a multimodal data source.

Description

Using Convolution network + Vision Transformer for 2D sketch-based 3D image retrieval.

Description

Using combined Knowledge Distillation + Lottery Ticket Hypothesis (LTH) + Channel Pruning for producing computer vision models with lower number of parameters.

Description

Developing fuzzy atrous convolutional layers for better image segmentation in medical datasets.

Description

Applying computer vision models on images to detect lanslides and generate information such as populated area, nearby roads, waterbodies, etc.

Description

Study of novel anti-spoof detection methods using textures and other image processing techniques.

Description

Forecasting of air pollution concentrations across measuring stations (spatial) given a history of pollution (temporal). Spatial layers use graph based structure simplified using Chebyshev's polynomial approximation while temporal layers use 1D CNN and a GLU layer. Multi-pollutant prediction simultaneously (multiple channels) is also explored.

Description

Developed during Google Summer of Code 2021 at Red Hen Lab. Detection of hand gestures using a deep learning model consisting of LSTM and CNN. Hand keypoints obtained from OpenPose.

Description

Idea presentation at MLDS 2022 conference for using transformers in single drone trajectory prediction system. Model design based on Anticipative Video Transformer. Paper: link

Description

Object Detection algorithms such as Mask R-CNN, YOLO, SSD are applied on a road images dataset collected using participatory sensing.

Description

Image Classification algorithms such as VGG, ResNet, Inception, Xception, DenseNet, SEResNet, and EfficientNet are applied on a road images dataset collected using participatory sensing.

Description

Created the backend for the webapp using Node. Problems solved like basic CRUD operations with MongoDB, providing shop-specified time slots to customers as per convenience and first-cum-first basis. Changing of slots is also allowed.

Description

Short introductory project on deep learning and tensorflow by implementing (with documentation) of a simple art generator using neural style transfer using a style image and a source image.