Stroke is a leading cause of motor disability. A majority of stroke survivors exhibit upper and lower limb motor impairments, ranging from incapability of reaching and grasping objects to limited ambulation. The objective of this project is to develop a personalized, community-based rehabilitation system to improve daily functions of stroke survivors. The system will include three essential components ������������������ a nanomaterial-enabled multifunctional wearable sensor network to monitor arm and leg functional activities; a low-power data acquisition, processing, and transmission protocol; and a user interface (i.e., smart phone APP) to communicate training outcomes to the users and clinicians and receive feedback from the users and clinicians. The proposed community-based rehabilitation system will enable personalized, continuous rehabilitation during daily activities.

Memory safety is essential. Despite decades of research, unauthorized memory reads and writes are still among the most common security attacks. The emerging persistent memory (PM) amplifies the importance of strong memory protections. As a promising supplement or substitute of DRAM as main memory, PM offers higher density, better scaling potential, lower idle power, and non-volatility, while retaining byte addressability and random accessibility. Data in a PMO is long lived; its existence and structure are preserved across process runs. The longevity, plus direct byte-addressability, makes it more vulnerable as attacks to a PMO could span across executions. This proposal aims to improve the understanding of the problem and provide innovative solutions to strengthen memory security for future NVM-based systems.

With the increasing penetration of behind-the-meter solar and energy storage, it is favored to leverage recent advances in artificial intelligence to enhance the accuracy of net-load forecasting, the observability of net-load variability, and the understanding of the coupling between net-load and demand response potentials. The proposed project will develop two models to address the hybrid probabilistic forecasting when small and large data sets are available. The first model will incorporate a new gradient boosting machine, in which a projection of the distribution into a Riemannian space is considered, whose corresponding natural gradient is expected to give better updates at each iteration than the state of the art. Meanwhile, a data-driven type-2 fuzzy system which generates monotone if-then rules will be developed to preprocess inputs. The second model consists of graph attention networks, transformers, and variational autoencoders. The graph attention networks overcome the theoretical issues with spectral based methods. The transformers ensure each time step to attend over all the time steps in the input sequence, compared with recurrent neural networks. The combination can give better spatiotemporal information. Moreover, those two models will be extended to forecast net-load with the consideration of demand response potentials, as a multi-target forecasting task.

The overarching goal of this proposal is to develop a generic HPC data registration and retrieval framework (named HPC-FAIR) to make both training data and AI models of scientific applications findable, accessible, interoperable, and reusable. This framework provisions significant speedup of the research and development of ML-based approaches for analyzing and optimizing scientific applications running on heterogeneous supercomputers. The datasets and AI models from HPC-FAIR will also serve as common baselines to quickly, consistently, and fairly evaluate new AI models for quality, complexity, and overhead.

Back-to-back failures of preseason hurricane predictions in the past two years left many hurricane forecasters wonder what can be done better to predict extreme hurricane activities. Although there have been many preseason hurricane prediction models, none of them is designed for predicting extreme activities which is especially challenging due to the imbalance between the high-dimensionality and large volume of data in the feature space and the relatively rare occurrences of such events. The team proposes a collection of novel explainable machine learning (ML) algorithms and implementations to address the challenges. These algorithms build on the team's prior work on high-performance and real-time machine learning and hurricane prediction.

This research proposes to advance the state of the art to holistic scalable autotuners, which tunes all levels of options for multiple optimization objectives at the same time. It will achieve this ambitious goal through the development of a set of novel techniques that efficiently handles the tremendous tuning space. These techniques take advantage of the synergies between all those options and goals by exploiting relevancy filtering (to quickly dispose of unhelpful options), locality of inference (that enables faster updates to out- dated tunings) and redundancy reduction (that reduces the search space for better tunings). This new autotuner will be a faster method for finding better tunings that satisfy more goals. To test this claim, this research will assess if this new tool can reduce the total computational resources required for effective SE data analytics by orders of magnitude.

The overall goal of this Phase 1 Convergence Accelerator (C-Accel) proposal is to develop what we know to be the first public-facing AI platform that assists individual workers and small employers with upskilling and career changes in a labor market increasingly characterized by automation, technological disruption, and AI recruiting. It will address key challenges faced by employees and employers in occupations most impacted by AI with labor market research, credential gap diagnostics, and support for job search and retraining in AI recruiting. Focusing on manufacturing in Phase I, we will develop and build support for an occupation predicted to lose about 20% jobs to automation by 2026, namely, machine operation hiring mostly male non-college workers. Exploring retraining resources, job search strategies in AI recruiting, and reemployment opportunities in related occupations requiring complementary skills, we aim to assist manufacturing workers with upskilling and retraining while developing educational materials to help prepare young generations for future jobs. Our innovative solution will be scaled up to a wide range of occupations and retraining programs in Phase II.

Modern supercomputers with heterogeneous components (e.g., GPUs) feature complex memory systems to meet the ever growing demands for data by processors. Putting data into the proper part of a memory system is essential for program performance, but is difficult to do. To address this challenge, we propose a new paradigm featuring three interacting components: 1) an extensible memory specification language to describe memory properties, 2) a compiler for analyzing data access patterns and transforming code for runtime adaptation, and 3) a data placement runtime to find and materialize the best data placements on the fly. The result will be a software framework (named XPlacer) that transforms OpenMP code to automatically place its data in memory in a way best suiting the GPU architecture, inputs, and program phases.

Contemporary architectures are adopting an integrated design of conventional CPUs with accelerators on the same die with access to the same memory, albeit with different coherence models. Examples include AMD's Fusion architecture, Intel's integrated main-stream CPU/GPU product line, and NVIDIA Tegra's integrated graphics processor family. Integrated GPUs feature shared caches and a common memory interconnect with multicore CPUs, which intensify resource contention in the memory hierarchy. This creates new challenges for data locality, task partitioning and scheduling, as well as program transformations. Most significantly, a program running on GPU warps and CPU cores may adversely affect performance and power of one another. The objective of this work is to understand these novel implications of fused architectures by studying their effects, qualifying their causes and quantifying the impacts on performance and energy efficiency. We propose to advance the state-of-the-art by creating spheres of isolation between CPU and GPU execution via novel systems mechanisms and compiler transformations that reduce cross-boundary contention with respect to shared hardware resources. This synergy between systems and compiler techniques has the potential to significantly improve performance and power guarantees for co-scheduling pgrams fragments on fused architectures. impact: The proposed work, if successful, has the potential to transform resource allocation and scheduling at the systems level and compiler optimizations at the program level to create a synergistic development environment with significant performance and power improvements and vastly increased isolation suitable for synergistic co-deployment of programs crossing boundaries on innovative fused architectures.

The proposed work for this project focuses on the development of machine learning training procedures and image post-processing and acquisition procedures. These methods will be applied to quantifying traits in agricultural applications and breeding.