Vishal Vinod

Relevant Coursework: Probabilistic Artificial Intelligence, Computational Photography, Unsupervised Learning, Virtual Reality Technology, Structured Prediction for NLP, Sampling and Reconstruction of Visual Appearance: From Denoising to View Synthesis, Deep Generative Models.

Relevant Courses: Artificial Intelligence, Machine Learning, Data Warehousing and Data Mining, Data Structures and Algorithms, Natural Language Processing

Developed a training free 3D face editing method for Generative NeRFs & obtained realistic 3D face edits for eyeglasses & age. Rendered dataset of 10K+ 3D faces from 2D images. Developed large-scale distributed ML training pipelines of upto 256 GPUs.

Investigated CRTF based sub-surface scattering in NeRFs for translucent objects with 10% improvement in novel view synthesis. Working on inverse rendering algorithms for physics-based face relighting with sub-surface scattering effects from 2D images.

Led development of TEGLO: a novel Generative NeRF for 3D reconstruction from single view images obtaining ≥74 db PSNR (≥200% improvement) over state of the art preserving fine details at megapixel resolutions [Under Review].

Simplified feature modifications in VW by implementing a reduction for data modification functions to register on the stack. Enabled on-the-fly feature manipulations for contextual bandit learning without requiring re-deploying the source.

Obtained ≥12% improvement over state-of-the-art in few-shot domain adaptation for RAW image enhancement in low light conditions for DSLR & smartphone cameras with new benchmark data [BMVC 2021, Best Student Paper Runner Up] Led the development of a self-supervised image translation method to translate daylight scenes to rain/fog/night and enabled a 5% improvement in image segmentation performance for self-driving cars in adverse weather [ICCV 2021, ILDAV]

Developed the first synthetic data generation method to enable generalization from handheld spectroradiometer data to drone based UAV cameras to bridge the high quality low-data problem - Climate Change project [JAG 2021]. Identified novel indices for Hyperspectral image analysis to estimate leaf water content and water stress early [ICCV 2021,CVPPA].

Implemented state-of-the-art ML models for TensorFlow mentored by Jaeyoun Kim, Hongkun Yu and Yanhui Liang from Google Brain.Led the development of FineGAN, a fine-grained image generation method with hierarchical disentanglement using TensorFlow. Investigated SIREN with implicit representation networks and evaluated the expressivity of Batchnorm for deep networks.

Implemented StackGAN: text-to-image Generative Adversarial Network (GAN) with two-stage image super-resolution using TF2. TF Model Garden: Built StackGAN, Mask R-CNN, Conditional VAE, and Face Aging with CycleGANs from the research papers.

Worked on "RainRoof: Automated Shared Rainwater Harvesting Prediction" under the guidance of Dr. Savita Choudhary (Springer LNDECT). Also worked on "Forecasting Dengue and Studying its Plausible Pandemy" under the guidance of Prof. Savita Choudhary (Elsevier). Worked on "Predicting the Network Structure using Harmonic Centrality Values in AODV Routing for Wireless Sensor Networks" under the guidance of Dr. Pallavi Venkatesh and Dr. Banu Prakash (SSN IJSTR 2019).

EEML summer school is focused on core topics regarding machine learning and artificial intelligence. The summer school includes both lectures and practical sessions (labs) to improve the theoretical and practical understanding of these topics. The school is aimed at graduate students. Participated in the Continual Learning track.

AI School to investigate and comprehend the limitations of the machine learning models and make them more fair, transparent and accountable. This school addresses some fundamental and pressing topics in fairness, accountability and transparency in AI.

Topics covered in the summer school broadly include comprehensive sessions on Machine Learning, AI, Knowledge Graphs, NLP, Game Theory, Discrete Mathematics, Linear Algebra, and Advanced Databases. Interactions with the professors was very engaging.