Bhagyashree Puranik

Multimodal content understanding at Prime Video

Robust machine learning: - Developed deep neural network architectures that are robust to out-of-distribution corruptions as well as mild adversarial perturbations by supplementing empirical risk minimization with layer-wise tilted exponential objectives motivated by increasing the signal-to-noise ratio at intermediate layers. -Modified the VGG and ResNet family of DNNs to introduce the TEXP models, which share attributes with transformers. Evaluations on CIFAR and ImageNet show significant gains in robustness compared to baselines. -Extensive evaluation by combining with SOTA data augmentation techniques and adversarial training show further gains. -In parallel, proposed an adversarial defense approach that is robust against perturbations, yielding performance competitive with existing minimax strategies, while providing a better robustness-accuracy trade-off for weaker attacks. -Skills: analytical design, extensive code development, working with complex codebases in PyTorch, thorough experimental evaluation, exposure to different vision models for classification and object detection. Fair machine learning: -Developed an MDP framework for sequential decision-making for dynamically influencing long-term fairness, when multiple agents are selecting resources from a common pool. -Proposed policies under de-centralized and centralized settings, showing the promise of positive feedback and studying robustness of policies under different evolution models for composition of the resource pool in a college admissions/hiring setting. -Skills: theoretical analysis, exposure to trustworthy ML, code development in Python and MATLAB.

Course: ECE130B - Signal Analysis

Course: ECE3 - Introduction to electrical engineering (Signal Processing and Communications)

Developing novel approaches to perform in-distribution and out-of-distribution generation through latent space manipulations in diffusion models.

Multimodal representation learning

- Designed a novel, low-cost, light-weight, deep learning approach to detect and localize humans inside car cabin and door open/close events using ultra-wideband sensors. - Deployed on a Nio car & showed its effectiveness in real-time. - Explored the use of text-driven generation of a sequence of animation images through StyleGAN like generative AI approach.

Summer'21 intern at the ADAS/Autonomy systems R&D team. Mentors: Urs Niesen & Meghana Bande. -Worked with the autonomy systems R&D team on building HD maps for autonomous driving from crowdsourced data. - Devised a novel spectral clustering and non-linear estimation based method to map traffic signs and lane boundaries, which was showcased at CES.

-Responsible for developing signal processing algorithms for communication applications. -Worked on the design, analysis and validation of IQ imbalance correction algorithm based on stochastic gradient descent for 5G base-station transceiver.

Worked with Prof P. Vijay Kumar on designing erasure codes for distributed storage and communication.

Undergrad project on an MMSE Framework for Projection Based Approaches to Interference Cancellation.

ME in Communication and Networks, Dept of ECE, IISc