Suggestions
Nikhil Naik
Lead Research Scientist at Salesforce Research. Hiring Full-time Researchers.
Professional Background
Nikhil Naik is a distinguished expert in the fields of Machine Learning and Computer Vision, currently contributing his extensive knowledge and skills as a Lead Research Scientist at Salesforce Research. With a robust commitment to advancing technological integration within various industries, Nikhil utilizes his background in economics to provide a unique perspective on data-driven solutions. His career trajectory encompasses significant roles at prestigious organizations, including positions as a Prize Postdoctoral Fellow at Harvard University and a Visiting Research Scientist at Google, where he continuously explores innovative applications of artificial intelligence.
Nikhil's tenure at Salesforce Research marks a significant phase in his professional journey, allowing him to lead cutting-edge projects centered around machine learning technologies that aim to revolutionize customer relationship management. His expertise enables him to direct research efforts that enhance the understanding of complex data interactions and improve the company’s mission to deliver intelligent solutions to its clients.
Prior to his current role, he honed his skills and expanded his knowledge base through fruitful internships at prominent tech corporations such as Microsoft Research and Samsung Research America. These experiences provided him with invaluable insights and practical skills critical to the development of high-impact technologies in machine learning and computer vision.
Education and Achievements
Nikhil Naik's impressive educational background lays a solid foundation for his achievements in the tech realm. He completed his Bachelor of Technology (B.Tech.) at the College of Engineering Pune, where he first cultivated his passion for engineering and technology. His academic journey continued at the Massachusetts Institute of Technology (MIT), where he earned both a Master of Science (M.S.) and subsequently a Doctor of Philosophy (Ph.D.). During his time at MIT, Nikhil dedicated himself to research that intertwines machine learning with real-world applications, significantly contributing to the academic community.
His educational achievements are complemented by a series of fellowships and research positions, including his time at Harvard University as a Prize Postdoctoral Fellow. This role not only advanced his research capabilities but also allowed him to contribute to a collaborative environment filled with leading scholars and innovators in various fields.
Achievements
Nikhil's contributions to the technology sector are profound and varied, with a noted focus on enabling businesses to harness the power of machine learning and computer vision. His research has resulted in significant advancements that have been recognized throughout the industry, especially in terms of enhancing efficiency and user experience in different consumer applications.
Additionally, his work has been published in leading scientific journals and presented at numerous conferences, showcasing his commitment to knowledge dissemination and the sharing of innovative ideas. Nikhil is known for his collaborative spirit, having partnered with experts across different disciplines to tackle complex problems, which has resulted in groundbreaking work that bridges the gap between theoretical research and practical application.
Through continuous learning and a passionate pursuit of innovation, Nikhil Naik remains a key figure in the intersection of technology and business strategy, shaping the future of machine learning and computer vision in significant and impactful ways.
Highlights
Excited to share our team’s latest work: Diffusion-DPO, a method to directly align image generation diffusion models to user preference. We finetuned the latest SDXL model with Diffusion-DPO. Our fine-tuned model obtains a 70% win-rate over SDXL in user studies, setting a new state-of-the-art for open source image generation models! More details here: arxiv.org/abs/2311.12908
Unlike large language models, there is a lack of scalable methods for aligning diffusion models to what people want. We were able to modify a recently developed LLM alignment method called Direct Preference Optimization (DPO) to enable direct finetuning of diffusion models to user preference data.
We used Diffusion-DPO to finetune the state-of-the-art Stable DIffusion XL model using 1 million pairs of user feedback data from the Pick-a-pic dataset. DPO-tuned SDXL is significantly better than SDXL in both visual appeal and fidelity to input prompt. In fact, the 3.5B parameter DPO-tuned SDXL-base model handily beats the much larger original 6.6B SDXL base + refinement model, removing the need to use the larger refinement model during inference.
It was great collaborating with Stefano Ermon’s group at the Stanford AI Lab for this work. Led by Bram Wallace with Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Caiming Xiong, Shafiq Joty and Stefano Ermon.
Introducing XGen-Image-1, our first foray into training large text-to-image-models. Trained for $75K using TPUs on the LAION dataset, XGen-Image-1 matches the performance of Stable Diffusion 1.5/2.1! Check out our blog to learn about our design decisions, training process, and key findings:
https://lnkd.in/gVxYeb5r
Work co-led with Bram Wallace, with valuable contributions from Akash Gokul Dongxu Li, Junnan Li and guidance from Caiming Xiong.
