![profile picture](https://dpl6hyzg28thp.cloudfront.net/media/93251acd-daca-4586-9bfe-02ff3cb42de4.jpg)
Vadim K
- Research Program Mentor
MS candidate at Brown University
Expertise
Deep Learning, AI/ML, NLP/LLMs, Gen-AI, Vision, Data Science, Graphics, Software Engineering, Computer Science
Bio
Hello! My name is Vadim! I primarily focus on computer science, and am most excited about topics in deep learning (especially generative models), computer vision and data science! I'm a prolific teaching assistant and just love helping out, so feel free to say hi! I also love anime and go to conventions quite often, so always happy to see other people into that stuff! Some key projects I've been a part of: - R package development projects (SimEd update 2.0, GEGravity). - NVIDIA DLI's Introduction to Robotic Simulations in Isaac Sim course. - Various deep learning projects and course development at Brown University. - Currently doing a bunch of LLM stuff at NVIDIA. Happy to talk about it, but not posting publicly yet.Project ideas
User-Guided Auto-Colorization
Deep learning models have had great success in generative contexts and have defined the state-of-the-art in many graphic-oriented tasks. One such task is auto-colorization. Given a training corpus, a model can be trained to automatically color images by learning to invert a simple greyscale function. However, even with an extremely large data corpus, models can still fail spectacularly with out-of-domain issues. Furthermore, creative endeavors may require specific color profiles which may be unrealistic or otherwise unlikely to be generated without special training. As such, human-supervised approaches have been of great interest and several projects have popped up to answer the call. In this project, we will explore another technique that is trained in tandem with color point hints which can be specified by the user (though are sampled randomly during training). We also use a data augmentation pipeline to scramble the color spectrum of the desired prediction, thus forcing the model to rely heavily on color hints to dictate its process.