Pranit Abhinav

Leveraged cloud platforms like GCP and AWS, conducted regular model training and benchmarking experiments. Utilized multiple GPUs with multiprocessing techniques to accelerate model training and optimize compute resource usage. Enhanced training efficiency through data and model parallelism in frameworks like TensorFlow and PyTorch. Developed and implemented an Mnasnet model with SSD head from scratch, achieving lightweight performance for passenger counting in vehicles. Led model training, fine-tuning, and benchmarking of advanced computer vision models, including DinoV2 and Detectron2 for object detection tasks. Preprocessed and managed large-scale datasets for deep learning pipelines using SQL and cloud buckets. Ran model benchmarking experiments on curated data sets, optimizing models through hyperparameter tuning and evaluating metrics such as accuracy, precision, recall, F1 score, ROC-AUC, and latency. Conducted a proof of concept for Neural Architecture Search on GCP Vertex AI with Google, optimizing model architecture and latency using search spaces like Spinenet and Mnasnet.

Developed and built a full peer-to-peer decentralized and distributed deep learning training system as a component of the Ravenverse framework to simultaneously update weights and communicate gradients among compute providers utilizing P2P servers. Developed and implemented techniques and architecture for splitting large functional AI models and their weights into smaller submodels and computational graphs to support distributed training on low compute resources. Demonstrated training large AI models using applications like neural style transfer model on Ravenverse for style transfer image generation. Worked on key functionalities necessary for the Ravenverse distribution framework. Demonstrated large AI models distributed training using imagenette dataset for architectures like resnet-50 and VGG.

Recognized and implemented many math operations required for Deep Learning computations in Requester and Provider libraries of Ravop and Ravpy. Worked on benchmarking contributors' devices using payloads for scheduling operations to a contributor based on its computing power. Contributed to Ravens centralized distribution server Ravsock for distributing ML tasks and operations. Documented several of Ravens distributed computing libraries including Ravop, Ravml, Ravdl, and RavJS. Implemented various machine learning algorithms, loss functions, and metric functions for Ravens ML libraries supporting distributed training.

Worked under Prof. Mohd Anul Haq as the sole undergraduate student on deep learning algorithms for GIS applications like hyperspectral analysis and groundwater forecasting using Azure compute credits. Trained and tuned ML models and presented findings as part of the AI for Earth conference.