Sony | R&D’s Post

Greetings, Diffusion models are a powerful generative framework, but come with expensive inference. Existing acceleration methods often compromise image quality or fail under complex conditioning. Now issue is being addressed by Meta researchers introducing Imagine Flash a novel distillation framework tailored to enable high-fidelity, diverse sample generation using just one to three steps. The three steps are: 1) Backward Distillation 2) Shifted Reconstruction Loss  3) Noise Correction Remarkably, it achieves performance comparable to the teacher model using only three denoising steps, enabling efficient high-quality generation. For more details check out: https://lnkd.in/dEq8CbYZ

Tutorial on Diffusion Models for Imaging and Vision 🎉🎉🎉 Stanley H. Chan Abstract The astonishing growth of generative tools in recent years has empowered many exciting applications in text-to-image generation and text-to-video generation. The underlying principle behind these generative tools is the concept of diffusion, a particular sampling mechanism that has overcome some shortcomings that were deemed difficult in the previous approaches. The goal of this tutorial is to discuss the essential ideas underlying the diffusion models. The target audience of this tutorial includes undergraduate and graduate students who are interested in doing research on diffusion models or applying these models to solve other problems. 👉 https://lnkd.in/dXjDQbDx #machinelearning

🔍 Are you eager to delve into the cutting-edge world of diffusion models in computer vision? Look no further! Check out this recent tutorial I've been following on the subject: https://lnkd.in/eFBhRBdV 💡 Understanding fundamental concepts is crucial in today's era of generative tools. This tutorial delves into the fascinating realm of diffusion models, shedding light on their underlying principles and their applications in text-to-image and text-to-video generation. 🎓 Targeted towards both undergraduate and graduate students, this tutorial provides an accessible yet comprehensive overview. With some dedication and effort, mastering these concepts can open doors to exciting research opportunities and real-world applications. 🚀 Don't miss out on this chance to explore the forefront of computer vision! Dive into the world of diffusion models and unleash your potential. #ComputerVision #GenerativeModels #DiffusionModels #ResearchOpportunity #Education #Students #Learning

Top important computer vision papers from week 9/10 to 15/10

"Tutorial on Diffusion Models for Imaging and Vision" by Stanley Chan "The astonishing growth of generative tools in recent years has empowered many exciting applications in text-to-image generation and text-to-video generation. The underlying principle behind these generative tools is the concept of diffusion, a particular sampling mechanism that has overcome some shortcomings that were deemed difficult in the previous approaches. The goal of this tutorial is to discuss the essential ideas underlying the diffusion models. The target audience of this tutorial includes undergraduate and graduate students who are interested in doing research on diffusion models or applying these models to solve other problems." Paper: https://lnkd.in/dYmgWg-y #diffusionmodels #computervision

Excited about Neural Fields and 3D Computer Vision? Check out this user-friendly DeepSDF repository: https://lnkd.in/ehxKzgrk Unlike other (excellent) resources, you can install it with a single command and use it for both 3D reconstruction and shape completion! I originally developed this tool for MSc students and early PhD students approaching the field. Hope you find this helpful :) Kudos to the DeepSDF authors: https://lnkd.in/ekDtquj5

Our paper "Waveformer for modelling dynamical systems" has been accepted for publication in the prestigious Mechanical Systems and Signal Processing! In this work, we introduce the Waveformer, a new operator learning approach for learning solution operators of dynamical systems. By leveraging wavelet transformation and transformers (and hence the name waveformer), our method excels in capturing both spatial multi-scale behavior and long horizon dynamics, addressing the shortcomings of existing neural operators. This work was primarily carried out by Navaneeth N, a Ph.D. scholar at #CSCCM #iitdelhi. A preprint version of the same can be accessed at the following link: https://lnkd.in/d5YG3quD The open access link (valid for 50 days) is available at the following link: https://lnkd.in/dA5Vh3Er #DynamicalSystems #MachineLearning #SciML #ScAI #amiitd

🌟 Exciting new research alert! Check out this pioneering work on a user authentication system based on human exhaled breath physics. This blog post explores the potential of using the structure of turbulence in exhaled breath as a unique biometric marker. Learn about the use of machine learning models and classical hypothesis testing in building user authentication algorithms. Read the full article here: https://bit.ly/48p14Vg [cs.CR] 📚 #innovation #biometrics #research

Some exciting news!! 💫Today I defended my research in Artificial Intelligence that I’ve been conducting for the past year at Utrecht University. Under the supervision of Chris Klink I worked on a Bayesian optimization algorithm for a simulation of a visual Brain Computer Interface designed to restore a rudimentary form of vision in blind patients. The procedure has been tested on 181 MRI scans from the Human Connectome Project, and the pipeline succeeds in virtually placing an arbitrary number of electrode arrays in the visual centers of the brain to maximize the generated artificial vision. The abstract and some visualizations of the results can be found on this dash app (runs best on a desktop or laptop): my-thesis-ai.appspot.com

I had the chance to publish a paper on a slight aside to the primary objectives of my research.   Firstly, it highlights how modern computing tools make it easy to calculate the shot noise SNR/sensitivity limit of an OCT system, from the raw measured data. Secondly, it also uses a numerical modelling technique to address something that has annoyed me for years. That being, the OCT community regularly does not correctly interpret USAF 1951 target images to define lateral resolution in an equivalent way to axial resolution. https://lnkd.in/e6hQpKkU