Heuristics for Auto-Tuning CNNs:
Developed heuristics for auto-tuning CNNs to improve the state-of-the-art model performance.
Recognized the necessity of extensive hyperparameter tuning, particularly for image pre-processing parameters (window size, batch size, downsampling rate) and the loss function ratio (Sorensen-Dice loss + alpha*Variational-Autoencoder Loss).
Designed experiments to optimize hyperparameters for the Brain Tumor Segmentation (BraTS2020) dataset. T
his extensible and reusable framework significantly streamlined the process, making it fully automated to achieve optimal tuning with minimal manual intervention.
Performance Improvements:
Enhanced the state-of-the-art model accuracy on the BraTS 2020 dataset by 5%, achieved with just one NVidia V100 32 GB GPU for the entire pipeline.
Experiments and Hyperparameter Tuning:
Set up a design of experiments for hyperparameter tuning, optimizing for the specific dataset and model. Emphasized the importance of window size, validated through response-surface ANOVA, which revealed that a wider cross-section and shallower depth were critical for improving accuracy, confirming correlations observed in existing literature.
Research Impact:
The developed framework is actively used in a collaboration between the medical imaging lab at IIT Madras and the Intel AI Lab in Chennai, India, for ongoing experiments on the BraTS2020 dataset.
