Walter Scheirer

🚀 Excited to share Compressive Light-Field Tokens (CLiFT) — a new scene representation for neural rendering developed by my students (Zhengqing Wang, Yuefan Wu, Jiacheng Chen, Fuyang Zhang). CLiFT is a Light-Field Token, compressed for compute-efficient and adaptive neural rendering, achieving (2x to 16x) reduction in storage and (66% to 100%) increase in FPS, while preserving high visual quality with only (0.54dB to 3dB) drop in PSNR. 🔗 Project: https://lnkd.in/gBqvMWi8 📦 Code coming 8/1!

166

3 Comments

There will be many basketball fans that think there are point defects this week. But for #FreePaperFriday from Purdue University School of Materials Engineering this week I'm going to highlight a paper about how MSE's look at point defects. In this recent topical review by Prof. Arun Mannodi Kanakkithodi and his student Habibur Rahman they cover ways that machine learning techniques can be used with first principles calculations to describe point defects in semiconductors. This is a busy space, but there are so many possible options (just an impurity isn't always the same, the location can make all the difference in the resulting properties) that an ML approach helps make the right choices for trying new chemistries and processing techniques. https://lnkd.in/gNhrYwtQ

108

2 Comments

Today, a friend shared a link with me to the Most Influential NeurIPS Papers summarized by Paper Digest: https://lnkd.in/eimhWXFB The ranking is based on citation counts. Our paper, "Learning Structured Sparsity in Deep Neural Networks," was ranked No. 10 (though I believe it is now at No. 9) among papers from NeurIPS 2016. This work (https://lnkd.in/eccRrG4S) was the first to demonstrate that deep neural networks need to be pruned structurally to achieve actual speedup on computing hardware. This approach later became a standard practice in neural network pruning in industry. The paper has been cited over 3,000 times since it was published.

242

3 Comments

Our work advancing quantum algorithms for nonlinear PDEs is now published in npj Quantum Information. This is an important potential application of quantum computers in simulating classical systems, but the nonlinear case is highly challenging for quantum algorithms. https://lnkd.in/gwURTaVA

175

5 Comments

Research in Focus: Purdue University School of Materials Engineering Can brittle intermetallics become deformable at room temperature? Researchers from Purdue MSE, led by Professors Xinghang Zhang and Haiyan Wang, have shown that they can, by introducing mobile dislocations across coherent interfaces. Their new publication in Scripta Materialia puts forward a CoAl/Fe nanolaminate system where mobile dislocations, initially generated in the Fe layers, continuously propagate into CoAl. This allows the intermetallic to exhibit high strength, significant work hardening, and most notably, plastic deformation at room temperature. This is an interesting finding for structural materials. The ability to retain the high strength of CoAl and overcome its intrinsic brittleness could make intermetallics viable for many engineering applications. This work provides information on the dislocation behavior at interfaces and also introduces a generalizable pathway for strengthening and toughening other brittle materials, including ceramics. The full article is available here: https://lnkd.in/gvVWg2kk Congratulations to Ke Xu, Anand Mathew, Zhongxia Shang, Yifan Zhang, Chao Shen, Benson Tsai, Jianan Shen, Haiyan Wang, and Xinghang Zhang on this impactful work. #MaterialsScience #Intermetallics #PurdueMSE #ResearchSpotlight #Nanomechanics #PurdueEngineering

91

#NVIDIAAcademicGrant Program is excited to announce new areas of focus for research proposals that 🧠 Enhance generative AI alignment and inference 🤖 Advance agentic model systems and reasoning ⚡ Push the boundaries in scientific simulation, quantum computing, and physics-informed machine learning 🔬 Develop foundation models in scientific domains like chemistry, climate, robotics Compute resources available to request include DGX Spark, RTX PRO 6000 Blackwell, and A100 GPU hours. 📅 Learn more about the proposal requirements and apply by June 30: https://bit.ly/3Hidwxm

95

1 Comment

This week Allen Zang, Xinan Chen, Eric Chitambar, Tian Zhong and myself published a new research article that provides guidance on how to optimize error management in quantum networks. The work prepared in collaboration with students and faculty at the Pritzker School of Molecular Engineering at the University of Chicago and the University of Illinois ECE department appeared in the Physical Review Letters as an Editors’ Suggestion: https://lnkd.in/gNsTrvyW Entanglement is required to enable quantum network applications such as distributed quantum computing, new secure communication protocols, and more precise metrology. However, errors and noise that degrade the quality of entanglement pose a significant challenge, necessitating the use of entanglement purification. Our work introduces the concept of universality, a property that guarantees that the quality of the purified entangled state is at least as good as the quality of the inputs. Our results show that it is fundamentally impossible to design a single one-size-fits-all universal protocol to counteract the noise. Instead, protocols must be tailored to the specifics of the system, such as the noise model. This work therefore carries real implications for the design of practical quantum communication networks. In our follow-up work, we have been investigating the impact of these realistic conditions and imperfections such as different quantum memory storage times and their decoherence properties: https://lnkd.in/g5triyJK. #QuantumNetworking #QuantumInformationScience #Entanglement

115

1 Comment

Imagine building a start-up on Deep Research system… without training a single model, and having the code open-sourced. 🤔 🚀 Excited to announce our new work from NVIDIA Research: Universal Deep Research (UDR): https://lnkd.in/gwRuU3Bw Deep research agents are one of the most widely used forms of AI agents. With UDR, we take a step forward: 🔍 Customizable strategy 🔛 research is not just about retrieving documents, but about shaping how the agent explores, verifies, and synthesizes knowledge. UDR shows that the research strategy itself can be adapted to the user’s goals. 🛠️ System-level behavior 🔛 the agent’s “deep research” workflow can be implemented without re-training or fine-tuning any LLMs. Instead, it’s orchestrated at the system level, making it lightweight, flexible, and easy to extend. 🌐 General-purpose design 🔛 UDR is universal: it can be applied across domains, whether for scientific discovery, technical due diligence, or business intelligence. AI agents are not just about better models - they are about better systems. We believe UDR opens the door to a new wave of startups and applications: anyone can now design a specialized “research assistant” that adapts to their domain and strategy. Learn more from products we offer in this domain: 🔗 AIQ Deep Research blueprint: https://lnkd.in/g2_4y3RR 🔗 NeMo Agent Toolkit: https://lnkd.in/gVPHTrzN Great project led by Peter Belcak.

126

3 Comments