Conducted ablation studies for exploring efficient heterogeneous (CPU + GPU) distributed training for language models such as BeRT and RoBeRTa over different factors such as batch size, number of CPU/GPU parallel workers etc. Ray was used for parallelizing CPU processes and Deepspeed’s ZeRO optimization was used for data parallelism along with mixed precision training for sentiment analysis.
Identified salient features in input images and generated adversarial masks using various techniques such as saliency gradient maps, GRAD-CAM and random patch masking. Created non-private representations of the input images using latent diffusion models, so that private information is not transmitted to downstream tasks such as FaceNet’s recognition model.
Tunable Visual prompting using data augmentation techniques such as DeepAugment, CutMix and CutOut for robust predictions for vision-language models like CLIP. Performance evaluations led to 2-3% improvement on different corruptions of CIFAR-C dataset