Search for: All records

We present Radius, a gradient sparsity algorithm and system to accelerate large foundation model (FM) training while preserving downstream task performance. Radius leverages two key insights in large FM pre-training: 1) only a small portion of gradients contribute to the model updates in each iteration, and 2) the spatial distribution of the gradients with large magnitude is stable over time. Radius overcomes the scaling problem of existing top-k sparsity methods, as it maintains the structure of sparse gradients thus avoids dense communication. We examine the convergence and speed of Radius on pre-training GPT models (355M and 2.0B) in data-parallel and compare it with the baseline top-k sparsification methods. Our results show that using the existing top-k method with AdamW optimizer fails to converge, and the training speed improvement with sparse communication is marginal. In contrast, Radius with 40% sparsity reduces per-step training time by 21% (19% for overall training time) across 64 NVIDIA A100 GPUs that are connected by the Slingshot 11 interconnect while preserving the downstream task performance. 

DNA has shown great biocompatibility, programmable mechanical properties, and precise structural addressabil- ity at the nanometer scale, rendering it a material for constructing versatile nanorobots for biomedical applica- tions. Here, we present the design principle, synthesis, and characterization of a DNA nanorobotic hand, called DNA NanoGripper, that contains a palm and four bendable fingers as inspired by naturally evolved human hands, bird claws, and bacteriophages. Each NanoGripper finger consists of three phalanges connected by three rotat- able joints that are bendable in response to the binding of other entities. NanoGripper functions are enabled and driven by the interactions between moieties attached to the fingers and their binding partners. We demonstrate that the NanoGripper can be engineered to effectively interact with and capture nanometer-scale objects, includ- ing gold nanoparticles, gold NanoUrchins, and SARS-CoV-2 virions. With multiple DNA aptamer nanoswitches programmed to generate a fluorescent signal that is enhanced on a photonic crystal platform, the NanoGripper functions as a highly sensitive biosensor that selectively detects intact SARS-CoV-2 virions in human saliva with a limit of detection of ~100 copies per milliliter, providing a sensitivity equal to that of reverse transcription quanti- tative polymerase chain reaction (RT-qPCR). Quantified by flow cytometry assays, we demonstrated that the NanoGripper-aptamer complex can effectively block viral entry into the host cells, suggesting its potential for in- hibiting virus infections. The design, synthesis, and characterization of a sophisticated nanomachine that can be tailored for specific applications highlight a promising pathway toward feasible and efficient solutions to the de- tection and potential inhibition of virus infections. 

There is an increasing attention to cellulose nanofibrils (CNFs) for food packaging applications due to their abundance, biodegradability, and low gas permeability. In this work, oxygen and water barrier performance is studied for bio-nanocomposite films formed by incorporation of two types of bentonite (PGN and PGV) at different loads (15, 30 and 45 wt%) into continuous CNF matrix. The resulting hybrid films were analyzed for their morphology, surface energy, mechanical strengths as well as water/oxygen barrier qualities. Both types of bentonite lowered the CNF degradation temperature and strength to some degree for reasons not so clear but perhaps due to partial disruption of the CNF H-bond network. It was revealed from microscopic study that clay particles form a layer within cellulose chains, resulting in alteration of composite structure. The contact angle analysis by polar and nonpolar liquids, suggested the PGN-containing samples were more hydrophilic; clay induced polar functionalities to the composite. While 15% PGN load reduced the water vapor transmission rate from 425 to 375 g/ m2 day, higher proportions of bentonite negatively affected this trend. Also, analysis of oxygen transmission rate showed the PGN effectively restricted the oxygen passage in dry state and to a lower extent at higher relative humidity. In WVTR analysis, PGN showed a superior performance over PGV attributable to its crystalline structure as evident in XRD patterns. The proposed hybrid CNF-BNT films in this study can present an eco-friendly alternative in packaging materials, especially where penetration of water vapor and oxygen is to be avoided. 

Abstract Physically based simulation is often combined with geometric mesh animation to add realistic soft‐body dynamics to virtual characters. This is commonly done using constraint‐based simulation whereby a soft‐tissue simulation is constrained to geometric animation of a subpart (or otherwise proxy representation) of the character. We observe that standard constraint‐based simulation suffers from an important flaw that limits the expressiveness of soft‐body dynamics. Namely, under correct physics, the frequency and amplitude of soft‐tissue dynamics arising from constraints (“inertial amplitude”) are coupled, and cannot be adjusted independently merely by adjusting the material properties of the model. This means that the space of physically based simulations is inherently limited and cannot capture all effects typically expected by computer animators. For example, animators need the ability to adjust the frequency, inertial amplitude, gravity sag and damping properties of the virtual character, independently from each other, as these are the primary visual characteristics of the soft‐tissue dynamics. We demonstrate that independence can be achieved by transforming the equations of motion into a non‐inertial reference coordinate frame, then scaling the resulting inertial forces, and then converting the equations of motion back to the inertial frame. Such scaling of inertia makes it possible for the animator to set the character's inertial amplitude independently from frequency. We also provide exact controls for the amount of character's gravity sag, and the damping properties. In our examples, we use linear blend skinning and pose‐space deformation for geometric mesh animation, and the Finite Element Method for soft‐body constrained simulation; but our idea of scaling inertial forces is general and applicable to other animation and simulation methods. We demonstrate our technique on several character examples. 

Recent advances in convolutional neural network (CNN) model interpretability have led to impressive progress in vi- sualizing and understanding model predictions. In partic- ular, gradient-based visual attention methods have driven much recent effort in using visual attention maps as a means for visual explanations. A key problem, however, is these methods are designed for classification and categorization tasks, and their extension to explaining generative models, e.g., variational autoencoders (VAE) is not trivial. In this work, we take a step towards bridging this crucial gap, proposing the first technique to visually explain VAEs by means of gradient-based attention. We present methods to generate visual attention from the learned latent space, and also demonstrate such attention explanations serve more than just explaining VAE predictions. We show how these attention maps can be used to localize anomalies in images, demonstrating state-of-the-art performance on the MVTec- AD dataset. We also show how they can be infused into model training, helping bootstrap the VAE into learning im- proved latent space disentanglement, demonstrated on the Dsprites dataset.