Sustainable Deep Learning & Frugal AI Framework
Cut CO₂ emissions by 42.6% with zero accuracy loss, and accelerated training by 33.7%.
Co-engineered a distributed model-parallel architecture for the Introvert pipeline using PyTorch RPC and Docker, reducing CO₂ emissions by 42.6% with zero accuracy degradation. Implemented an optimization pipeline featuring mixed-precision training (AMP), torch.compile and asynchronous data loading, accelerating training time by 33.7%. Formulated a composite frugality metric based on the Analytic Hierarchy Process (AHP) to evaluate AI deployments across energy, memory, FLOPs and predictive accuracy together.
In plain terms
Training large AI models burns a lot of electricity. This work splits the training across machines and tunes how it runs, so the same model comes out just as accurate for far less energy.
