Recent advances in Dynamic Sparse Training (DST) have pushed the frontier of sparse neural network training in structured and unstructured contexts, matching dense-model performance while drastically reducing parameter counts to facilitate model scaling. However, unstructured sparsity often fails to translate into practical speedups on modern hardware. To address this shortcoming, we propose DynaDiag, a novel structured sparse-to-sparse DST method that performs at par with unstructured sparsity. DynaDiag enforces a diagonal sparsity pattern throughout training and preserves sparse computation in forward and backward passes. We further leverage the diagonal structure to accelerate computation via a custom CUDA kernel, rendering the method hardware-friendly. Empirical evaluations on diverse neural architectures demonstrate that our method maintains accuracy on par with unstructured counterparts while benefiting from tangible computational gains. Notably, with 90\% sparse linear layers in ViTs, we observe up to a 3.13x speedup in online inference without sacrificing model performance and a 1.59x speedup in training on a GPU compared to equivalent unstructured layers.

Visual servoing for the development of autonomous robotic systems capable of administering UltraSound (US) guided regional anesthesia requires real-time segmentation of nerves, needle tip localization and needle trajectory extrapolation. First, we recruited 227 patients to build a large dataset of 41,000 anesthesiologist annotated images from US videos of brachial plexus nerves and developed models to localize nerves in the US images. Generalizability of the best suited model was tested on the datasets constructed from separate US scanners. Using these nerve segmentation predictions, we define automated anesthesia needle targets by fitting an ellipse to the nerve contours. Next, we developed an image analysis tool to guide the needle toward their targets. For the segmentation of the needle, a natural RGB pre-trained neural network was first fine-tuned on a large US dataset for domain transfer and then adapted for the needle using a small dataset. The segmented needle’s trajectory angle is calculated using Radon transformation and the trajectory is extrapolated from the needle tip. The intersection of the extrapolated trajectory with the needle target guides the needle navigation for drug delivery. The needle trajectory’s average error was within acceptable range of 5 mm as per experienced anesthesiologists. The entire dataset has been released publicly for further study by the research community at https://github.com/Regional-US/

As Neural Processing Units (NPU) or accelerators are increasingly deployed in a variety of applications including safety critical applications such as autonomous vehicle, and medical imaging, it is critical to understand the fault-tolerance nature of the NPUs. We present a reliability study of Arm's Ethos-U55, an important industrial-scale NPU being utilised in embedded and IoT applications. We perform large scale RTL-level fault injections to characterize Ethos-U55 against the Automotive Safety Integrity Level D (ASIL-D) resiliency standard commonly used for safety-critical applications such as autonomous vehicles. We show that, under soft errors, all four configurations of the NPU fall short of the required level of resiliency for a variety of neural networks running on the NPU. We show that it is possible to meet the ASIL-D level resiliency without resorting to conventional strategies like Dual Core Lock Step (DCLS) that has an area overhead of 100%. We achieve so through selective protection, where hardware structures are selectively protected (e.g., duplicated, hardened) based on their sensitivity to soft errors and their silicon areas. To identify the optimal configuration that minimizes the area overhead while meeting the ASIL-D standard, the main challenge is the large search space associated with the time-consuming RTL simulation. To address this challenge, we present a statistical analysis tool that is validated against Arm silicon and that allows us to quickly navigate hundreds of billions of fault sites without exhaustive RTL fault injections. We show that by carefully duplicating a small fraction of the functional blocks and hardening the Flops in other blocks meets the ASIL-D safety standard while introducing an area overhead of only 38 %.

Battery life is an increasingly urgent challenge for today's untethered VR and AR devices. However, the power efficiency of head-mounted displays is naturally at odds with growing computational requirements driven by better resolution, refresh rate, and dynamic ranges, all of which reduce the sustained usage time of untethered AR/VR devices. For instance, the Oculus Quest 2, under a fully-charged battery, can sustain only 2 to 3 hours of operation time. Prior display power reduction techniques mostly target smartphone displays. Directly applying smartphone display power reduction techniques, however, degrades the visual perception in AR/VR with noticeable artifacts. For instance, the "power-saving mode" on smartphones uniformly lowers the pixel luminance across the display and, as a result, presents an overall darkened visual perception to users if directly applied to VR content. Our key insight is that VR display power reduction must be cognizant of the gaze-contingent nature of high field-of-view VR displays. To that end, we present a gaze-contingent system that, without degrading luminance, minimizes the display power consumption while preserving high visual fidelity when users actively view immersive video sequences. This is enabled by constructing 1) a gaze-contingent color discrimination model through psychophysical studies, and 2) a display power model (with respect to pixel color) through real-device measurements. Critically, due to the careful design decisions made in constructing the two models, our algorithm is cast as a constrained optimization problem with a closed-form solution, which can be implemented as a real-time, image-space shader. We evaluate our system using a series of psychophysical studies and large-scale analyses on natural images. Experiment results show that our system reduces the display power by as much as 24% (14% on average) with little to no perceptual fidelity degradation.

As throughput of neural network accelerator datapaths have grown, memory has consistently fallen behind. Although attempts have been made to improve performance of neural networks via approaches such as batching, several layers often starve for memory bandwidth when used for tasks such as online inferencing. In order to mitigate memory bandwidth limitations, we propose a near memory accelerator for mobile devices based on Transport-Triggered Architecture (TTA) and evaluate its performance benefits compared to the existing approaches. Through experiments we demonstrate that our proposed accelerator achieves up to 4.3× speedup with respect to a non-near memory accelerator and the proposed TTA data-path achieves area and energy efficiency close to fixed-function datapaths while offering programmability similar to VLIW cores.

In past few years we have observed an increase in the usage of RGBD sensors in mobile devices. These sensors provide a good estimate of the depth map for the camera frame, which can be used in numerous augmented reality applications. This paper presents a new visual inertial odometry (VIO) system, which uses measurements from a RGBD sensor and an inertial measurement unit (IMU) sensor for estimating the motion state of the mobile device. The resulting system is called the depth-aided VIO (DVIO) system. In this system we add the depth measurement as part of the nonlinear optimization process. Specifically, we propose methods to use the depth measurement using one-dimensional (1D) feature parameterization as well as three-dimensional (3D) feature parameterization. In addition, we propose to utilize the depth measurement for estimating time offset between the unsynchronized IMU and the RGBD sensors. Last but not least, we propose a novel block-based marginalization approach to speed up the marginalization processes and maintain the real-time performance of the overall system. Experimental results validate that the proposed DVIO system outperforms the other state-of-the-art VIO systems in terms of trajectory accuracy as well as processing time.

In recent times, there is a paradigm shift to cloud services that offer on-demand computer system resources, especially data storage and computing power. The main reason for the shift is that it removes the user's active participation to perform computationally intensive tasks. However, current cloud-based services incur high user latency as being deployed very far from the user. One alternative solution to the traditional cloud-based paradigm is drone-based edge computing. In drone edge computing, drones are located near the user and deployed to provide data offload services. There have been many works that have addressed the issue of efficient task assignment in edge devices. This paper presents a concrete analytical performance model for drone cloudlet networks and factors that influence the service response time to the user. The results can be helpful for network administrators to make the current edge computing paradigm faster, more robust and, cost-effective.