OBJECTIVES: Electronic Health Records (EHRs) sampled from different populations can introduce unwanted biases, limit individual-level data sharing, and make the data and fitted model hardly transferable across different population groups. In this context, our main goal is to design an effective method to transfer knowledge between population groups, with computable guarantees for suitability, and that can be applied to quantify treatment disparities. MATERIALS AND METHODS: For a model trained in an embedded feature space of one subgroup, our proposed framework, Optimal Transport-based Transfer Learning for EHRs (OTTEHR), combines feature embedding of the data and unbalanced optimal transport (OT) for domain adaptation to another population group. To test our method, we processed and divided the MIMIC-III and MIMIC-IV databases into multiple population groups using ICD codes and multiple labels. RESULTS: We derive a theoretical bound for the generalization error of our method, and interpret it in terms of the Wasserstein distance, unbalancedness between the source and target domains, and labeling divergence, which can be used as a guide for assessing the suitability of binary classification and regression tasks. In general, our method achieves better accuracy and computational efficiency compared with standard and machine learning transfer learning methods on various tasks. Upon testing our method for populations with different insurance plans, we detect various levels of disparities in hospital duration stay between groups. DISCUSSION AND CONCLUSION: By leveraging tools from OT theory, our proposed framework allows to compare statistical models on EHR data between different population groups. As a potential application for clinical decision making, we quantify treatment disparities between different population groups. Future directions include applying OTTEHR to broader regression and classification tasks and extending the method to semi-supervised learning.

Intermittently-operating embedded computing platforms powered by energy harvesting must frequently checkpoint their computation state. Using non-volatile memory reduces checkpoint size by eliminating the need to checkpoint volatile memory but increases checkpoint frequency to cover Write After Read (WAR) dependencies. Additionally, non-volatile memory is significantly slower to access - while consuming more energy than its volatile counterpart - suggesting the use of a data cache. Unfortunately, existing data cache solutions do not fit the challenges of intermittent computing and often require additional hardware or software to detect WARs. In this paper, we extend the data cache by integrating it with WAR detection - dropping the need for an additional memory tracker. This idea forms the basis of NACHO: a data cache tailored to intermittent computing. NACHO, on average, reduces intermittent computing runtime overhead by 54% compared to state of the art cache-based systems. It also reduces the number of non-volatile memory writes by 82% compared to a data cache-less system, and 18% on average compared to multiple state of the art cache-based systems.

We present hardware/software techniques to intelligently regulate supply voltage and clock frequency of intermittently computing devices. These devices rely on ambient energy harvesting to power their operation and small capacitors as energy buffers. Statically setting their clock frequency fails to capture the unique relations these devices expose between capacitor voltage, energy efficiency at a given operating frequency, and the corresponding operating range. Existing dynamic voltage and frequency scaling techniques are also largely inapplicable due to extreme energy scarcity and peculiar hardware features. We introduce two hardware/software co-designs that accommodate the distinct hardware features and function within a constrained energy envelope, offering varied tradeoffs and functionalities. Our experimental evaluation combines tests on custom-manufactured hardware and detailed emulation experiments. The data gathered indicate that our approaches result in up to 3.75× reduced energy consumption and 12× swifter execution times compared to the considered baselines, all while utilizing smaller capacitors to accomplish identical workloads.

Continuous health monitoring is crucial to ensuring better health and taking preventive measures just-in-time. Existing battery-powered health wearables pose a significant limitation to continuous monitoring as batteries wear out after fixed energy cycles and need replacement. Ambient energy harvesting unlocks battery-free sensing but it suffers from spatio-temporal variability, making it unfit for health sensing. Intra-body power transfer (IBPT) provides an alternative energy source for battery-free operation, however, it can only provide limited energy in order to ensure wearer's safety. Existing system support is designed to maximize computational progress in a single energy cycle, thus wasting energy on computations that become stale in the next energy cycle. We instantiate an IBPT-powered health wearable capable of supporting multiple health sensors. To cope with lower incoming energy, we introduce BodyOS; a system support that exposes programming constructs for domain experts to express health applications in terms of the inherent dependencies of bio-signals being monitored by the application. By avoiding unnecessary sensing operations, BodyOS allows energy-efficient application execution and faster capacitor recharge while ensuring that the data sensed by the application is always useful. We evaluate BodyOS to show that it significantly improves energy efficiency, thus increasing the on-time and number of data points collected by the device.

The rise of 5G mobile networks necessitated virtualized Radio Access Network (vRAN) deployment, whereby radio functions are softwarized. This involves disaggregating traditional Baseband Units (BBUs) into Radio units (RU), Distributed Units (DUs), and Centralized Units (CUs) and dynamically instantiating radio network functions onto these for efficient resource optimization. However, establishing a robust vRAN for 5 G encounters challenges due to diverse application scenarios like Ultra-Reliable Low-Latency Communication (URLLC), Enhanced Mobile Broadband (eMBB), and Massive Machine Type Communication (mMTC), each with a specific service requirement. This paper studies the challenges involved in the optimal placement of radio network functions onto the disaggregated vRAN components while slicing the network to support diverse 5 G application scenarios. The paper models a 5 G vRAN and proposes an Integer Linear Programming (ILP) framework - SliAvailRAN, which maximizes the service provider’s profit by optimally allocating virtualized network functions for 5G application scenarios. We evaluate SliAvailRAN on two realistic topologies (with multiple variants) and show results for request acceptance and achieved centralization levels.

Task-based intermittent software systems always re-execute peripheral input/output (I/O) operations upon power failures since tasks have all-or-nothing semantics. Re-executed I/O wastes significant time and energy and risks memory inconsistency. This paper presents EaseIO, a new task-based intermittent system that remedies these problems. EaseIO programming interface introduces re-execution semantics for I/O operations to facilitate safe and efficient I/O management for intermittent applications. EaseIO compiler front-end considers the programmer-annotated I/O re-execution semantics to preserve the task's energy efficiency and idem-potency. EaseIO runtime introduces regional privatization to eliminate memory inconsistency caused by idempotence bugs. Our evaluation shows that EaseIO reduces the wasted useful I/O work by up to 3× and total execution time by up to 44% by avoiding 76% of the redundant I/O operations, as compared to the state-of-the-art approaches for intermittent computing. Moreover, for the first time, EaseIO ensures memory consistency during DMA-based I/O operations.

Battery-free and intermittently powered devices offer long lifetimes and enable deployment in new applications and environments. Unfortunately, developing sophisticated inference-capable applications is still challenging due to the lack of platform support for more advanced (32-bit) microprocessors and specialized accelerators---which can execute data-intensive machine learning tasks, but add complexity across the stack when dealing with intermittent power. We present Protean to bridge the platform gap for inference-capable battery-free sensors. Designed for runtime scalability, meeting the dynamic range of energy harvesters with matching heterogeneous processing elements like neural network accelerators. We develop a modular "plug-and-play" hardware platform, SuperSensor, with a reconfigurable energy storage circuit that powers a 32-bit ARM-based microcontroller with a convolutional neural network accelerator. An adaptive task-based runtime system, Chameleon, provides intermittency-proof execution of machine learning tasks across heterogeneous processing elements. The runtime automatically scales and dispatches these tasks based on incoming energy, current state, and programmer annotations. A code generator, Metamorph, automates conversion of ML models to intermittent safe execution across heterogeneous compute elements. We evaluate Protean with audio and image workloads and demonstrate up to 666x improvement in inference energy efficiency by enabling usage of modern computational elements within intermittent computing. Further, Protean provides up to 166% higher throughput compared to non-adaptive baselines.

In developing countries, majority of the households use overhead water tanks to have running water. These water tanks are exposed to the elements, which usually render the tap water uncomfortable to use, given the extreme subtropical weather conditions. Externally weatherproofing these tanks to maintain the groundwater temperature is short-lived, and only results in a marginal (0.5°C–1°C) improvement in tap water temperature. We propose Ashray , an IoT-inspired, intelligent system to minimize the exposure of water to the elements thereby maintaining its temperature close to that of the groundwater. Ashray learns the water demand patterns of a household and pumps water into the overhead tank only when necessary. The predictive, machine learning based, approach of Ashray improves water comfort by up to 8°C in summers and 3°C in winters, on average. Ashray is retrofitted into existing infrastructure with a hardware prototyping cost of $27, whereas it can save up to 16% on water heating costs, through reduction in natural gas consumption, by leveraging groundwater temperature. Moreover, we also consider a transiently-powered Ashray , which uses the energy harvested from the ambient environment, and propose an intermittent data pipeline to improve its prediction accuracy. The transiently-powered Ashray is suitable for long-term deployment, requires minimal maintenance and delivers approximately the same performance. Ashray has the potential to improve the thermal comfort and reduce energy costs for millions of households in developing countries.

In developing countries, majority of the households use overhead water tanks to have running water in their taps. These water tanks are exposed to the elements, which usually render the tap water uncomfortable to use, given the extreme subtropical weather conditions. Externally weatherproofing these tanks to maintain the groundwater temperature is short-lived, and only results in a marginal (0.5 −1◦C) improvement in tap water temperature. We propose Ashray, an IoT-inspired, intelligent system to minimize the exposure of water to the elements thereby maintaining its temperature close to that of the groundwater. Ashray learns the water demand patterns of a household and pumps water into the overhead tank only when necessary. The predictive, machine learning based, approach of Ashray improves water comfort by up to 8◦C in summers and 3◦C in winters, on average. Ashray is retrofitted into existing infrastructure with a hardware prototyping cost of $27, whereas it can save up to 16% on water heating costs, through reduction in natural gas consumption, by leveraging groundwater temperature. Our proposed system, Ashray, can positively impact the lives of millions of people in developing countries.

Intermittently powered embedded devices ensure forward progress of programs through state checkpointing in non-volatile memory. Checkpointing is, however, expensive in energy and adds to the execution times. To minimize this overhead, we present DICE, a system that renders differential checkpointing profitable on these devices. DICE is unique because it is a software-only technique and efficient because it only operates in volatile main memory to evaluate the differential. DICE may be integrated with reactive (Hibernus) or proactive (MementOS, HarvOS) checkpointing systems, and arbitrary code can be enabled with DICE using automatic code-instrumentation requiring no additional programmer effort. By reducing the cost of checkpoints, DICE cuts the peak energy demand of these devices, allowing operation with energy buffers that are one-eighth of the size originally required, thus leading to benefits such as smaller device footprints and faster recharging to operational voltage level. The impact on final performance is striking: with DICE, Hibernus requires one order of magnitude fewer checkpoints and one order of magnitude shorter time to complete a workload in real-world settings.

Transiently powered computers (TPCs) form the foundation of the battery-less Internet of Things, using energy harvesting and small capacitors to power their operation. This kind of power supply is characterized by extreme variations in supply voltage, as capacitors charge when harvesting energy and discharge when computing. We experimentally find that these variations cause marked fluctuations in clock speed and power consumption . Such a deceptively minor observation is overlooked in existing literature. Systems are thus designed and parameterized in overly conservative ways, missing on a number of optimizations. We rather demonstrate that it is possible to accurately model and concretely capitalize on these fluctuations. We derive an energy model as a function of supply voltage and prove its use in two settings. First, we develop EPIC, a compile-time energy analysis tool. We use it to substitute for the constant power assumption in existing analysis techniques, giving programmers accurate information on worst-case energy consumption of programs. When using EPIC with existing TPC system support, run-time energy efficiency drastically improves, eventually leading up to a 350% speedup in the time to complete a fixed workload. Further, when using EPIC with existing debugging tools, it avoids unnecessary program changes that hurt energy efficiency. Next, we extend the MSPsim emulator and explore its use in parameterizing a different TPC system support. The improvements in energy efficiency yield up to more than 1000% time speedup to complete a fixed workload.

Embedded devices running on ambient energy perform computations intermittently, depending upon energy availability. System support ensures forward progress of programs through state checkpointing in non-volatile memory. Checkpointing is, however, expensive in energy and adds to execution times. To reduce this overhead, we present DICE, a system design that efficiently achieves differential checkpointing in intermittent computing. Distinctive traits of DICE are its software-only nature and its ability to only operate in volatile main memory to determine differentials. DICE works with arbitrary programs using automatic code instrumentation, thus requiring no programmer intervention, and can be integrated with both reactive (Hibernus) or proactive (MementOS, HarvOS) checkpointing systems. By reducing the cost of checkpoints, performance markedly improves. For example, using DICE, Hibernus requires one order of magnitude shorter time to complete a fixed workload in real-world settings.

Transiently-powered computers (TPCs) lay the basis for a battery-less Internet of Things, using energy harvesting and small capacitors to power their operation. This power supply is characterized by extreme variations in supply voltage, as capacitors charge when harvesting energy and discharge when computing. We experimentally find that these variations cause marked fluctuations in clock speed and power consumption, which determine energy efficiency. We demonstrate that it is possible to accurately model and concretely capitalize on these fluctuations. We derive an energy model as a function of supply voltage and develop EPIC, a compile-time energy analysis tool. We use EPIC to substitute for the constant power assumption in existing analysis techniques, giving programmers accurate information on worst-case energy consumption of programs. When using EPIC with existing TPC system support, run-time energy efficiency drastically improves, eventually leading up to a 350% speedup in the time to complete a fixed workload. Further, when using EPIC with existing debugging tools, programmers avoid unnecessary program changes that hurt energy efficiency.

We propose a set of differential techniques to allow transientlypowered embedded devices reduce the amount of data written on non-volatile memory during checkpoints used to cross times of energy unavailability. These techniques track modifications in the application state to isolate data from the slice of the previous checkpoint that remains unaltered. At the following checkpoint, our approach may thus only update the parts it detects as modified. This allows us to shift part of the energy budget from checkpointing overhead to useful computations, yielding better overall energy efficiency.

We propose incremental checkpointing techniques enabling transiently powered devices to retain computational state across multiple activation cycles. As opposed to the existing approaches, which checkpoint complete program state, the proposed techniques keep track of modified RAM locations to incrementally update the retained state in secondary memory, significantly reducing checkpointing overhead both in terms of time and energy.

This paper evaluates the tele-operation of a mobile sensor platform. In current operations, this land vehicle is controlled by a team consisting of a driver and a sensor operator. Our experiment is the first attempt, for this system, to assess the impact, if any, of using a single operator versus a team of two operators, in order to inform staffing decisions as well as the design of future automated functions. The experiment was conducted using a simulator and 24 participants: 8 single operators and 8 teams of two operators. Lower mission completion times arose for the two-operator condition in spite of any extra time and workload generated by communication required. Operators in teams also assessed the system as more usable. These findings give support to the team strategy for tele-operating mobile sensor platforms, and have implications for the staffing and design of similarly complex uninhabited vehicle systems.