Yash Vardhan Sharma Email & Phone Number

Bengaluru, Ka, In

Bengaluru, Karnataka, India

Engaged in diverse on-device ML initiatives, specializing in federated learning (FL), hyper-personalization, and low-latency model execution to optimize cloud costs.Researched extensively in on-device ML literature, focusing on FL, hyper-personalization via fine-tuning, and cohort-wise clustering techniques.Key development contributions include FL and hyper-personalization frameworks for in-house simulations, implementing model pipelines for ONNX-Runtime, TF-lite, and MNN, and… Show more Engaged in diverse on-device ML initiatives, specializing in federated learning (FL), hyper-personalization, and low-latency model execution to optimize cloud costs.Researched extensively in on-device ML literature, focusing on FL, hyper-personalization via fine-tuning, and cohort-wise clustering techniques.Key development contributions include FL and hyper-personalization frameworks for in-house simulations, implementing model pipelines for ONNX-Runtime, TF-lite, and MNN, and creating scripts for on-device ML model pre/post-processing.Developed a comprehensive mobile testing framework for large-scale on-device ML execution, ensuring prediction accuracy, latency optimization, and memory management. This involved significant input into both the Android SDK and the Python driver orchestrating the tests. Show less

Worked on the problem of bias and non-robustness of neural nets in computer vision, supervised by Prof Vinay Namboodiri.Developed novel Gradient Based Activations in the adversarial training setup of de-bias to achieve state-of-the-art results on current benchmarks.Above research got accepted at AAAI'22 (an A* ML Conference)

- Organised and presented in reading group discussions on the ML research papers.- Participated in Inter-IIT tech meet in the traffic sign recognition event by bosh, bagged silver medal.- Submitted report in ML-reproducibility challenge on a paper on uncertainty quantification, added a novel extension on it to improve on the results. Report was appreciated by the authors of the paper.

- Implemented G3PCX, DE, PSO, CMA-ES optimization algorithms.- Developed island versions using distributed computation to increase computational efficiency.- Developed Surrogate ANNs for modelling simulations to increase computational efficiency for expensive simulations- Achieved State-of-the-art performance on various benchmark problems