Timely decision making is critical to the effectiveness of mobile health (mHealth) interventions. At predefined timepoints called "decision points," intelligent mHealth systems such as just-in-time adaptive interventions (JITAIs) estimate an individual's biobehavioral context from sensor or survey data and determine whether and how to intervene. For interventions targeting habitual behavior (e.g., oral hygiene), effectiveness often hinges on delivering support shortly before the target behavior is likely to occur. Current practice schedules decision points at a fixed interval (e.g., one hour) before user-provided behavior times, and the fixed interval is kept the same for all individuals. However, this one-size-fits-all approach performs poorly for individuals with irregular routines, often scheduling decision points after the target behavior has already occurred, rendering interventions ineffective. In this paper, we propose SigmaScheduling, a method to dynamically schedule decision points based on uncertainty in predicted behavior times. When behavior timing is more predictable, SigmaScheduling schedules decision points closer to the predicted behavior time; when timing is less certain, SigmaScheduling schedules decision points earlier, increasing the likelihood of timely intervention. We evaluated SigmaScheduling using real-world data from 68 participants in a 10-week trial of Oralytics, a JITAI designed to improve daily toothbrushing. SigmaScheduling increased the likelihood that decision points preceded brushing events in at least 70% of cases, preserving opportunities to intervene and impact behavior. Our results indicate that SigmaScheduling can advance precision mHealth, particularly for JITAIs targeting time-sensitive, habitual behaviors such as oral hygiene or dietary habits.

In this work, we develop an end-to-end differentiable multi-view visual tracking architecture and explore fine-tuning model parameters via gradient-based optimization and automatic differentiation. We consider a setting with multiple camera nodes distributed in the tracking environment that collaboratively track objects. The architecture that we construct includes within-image-plane deep learning-based detection models, probabilistic camera models, object dynamics models, and an$N$-object Kalman filter-based tracking model. We demonstrate fully differentiable choices for each of these components, enabling learning and fine-tuning of the parameters of all system components based on different forms of supervision. Our results show performance gains for$N$-object tracking when fine-tuning the parameters of the system for end-to-end tracking performance.

The success of self-supervised contrastive learning hinges on identifying positive data pairs, such that when they are pushed together in embedding space, the space encodes useful information for subsequent downstream tasks. Constructing positive pairs is non-trivial as the pairing must be similar enough to reflect a shared semantic meaning, but different enough to capture within-class variation. Classical approaches in vision use augmentations to exploit well-established invariances to construct positive pairs, but invariances in the time-series domain are much less obvious. In our work, we propose a novel method of using a learned measure for identifying positive pairs. Our Retrieval-Based Reconstruction (REBAR) measure measures the similarity between two sequences as the reconstruction error that results from reconstructing one sequence with retrieved information from the other. Then, if the two sequences have high REBAR similarity, we label them as a positive pair. Through validation experiments, we show that the REBAR error is a predictor of mutual class membership. Once integrated into a contrastive learning framework, our REBAR method learns an embedding that achieves state-of-the-art performance on downstream tasks across various modalities.

Localization is a critical technology for various applications ranging from navigation and surveillance to assisted living. Localization systems typically fuse information from sensors viewing the scene from different perspectives to estimate the target location while also employing multiple modalities for enhanced robustness and accuracy. Recently, such systems have employed end-to-end deep neural models trained on large datasets due to their superior performance and ability to handle data from diverse sensor modalities. However, such neural models are often trained on data collected from a particular set of sensor poses (i.e., locations and orientations). During real-world deployments, slight deviations from these sensor poses can result in extreme inaccuracies. To address this challenge, we introduce FlexLoc, which employs conditional neural networks to inject node perspective information to adapt the localization pipeline. Specifically, a small subset of model weights are derived from node poses at run time, enabling accurate generalization to unseen perspectives with minimal additional overhead. Our evaluations on a multimodal, multi-view indoor tracking dataset showcase that FlexLoc improves the localization accuracy by almost 50% in the zero-shot case (no calibration data available) compared to the baselines. The source code of FlexLoc is available in https://github.com/nesl/FlexLoc.

OBJECTIVE: Digital behavior change interventions (DBCIs) are feasibly effective tools for addressing physical activity. However, in-depth understanding of participants' long-term engagement with DBCIs remains sparse. Since the effectiveness of DBCIs to impact behavior change depends, in part, upon participant engagement, there is a need to better understand engagement as a dynamic process in response to an individual's ever-changing biological, psychological, social, and environmental context. METHODS: The year-long micro-randomized trial (MRT) HeartSteps II provides an unprecedented opportunity to investigate DBCI engagement among ethnically diverse participants. We combined data streams from wearable sensors (Fitbit Versa, i.e., walking behavior), the HeartSteps II app (i.e. page views), and ecological momentary assessments (EMAs, i.e. perceived intrinsic and extrinsic motivation) to build the idiographic models. A system identification approach and a fluid analogy model were used to conduct autoregressive with exogenous input (ARX) analyses that tested hypothesized relationships between these variables inspired by Self-Determination Theory (SDT) with DBCI engagement through time. RESULTS: Data from 11 HeartSteps II participants was used to test aspects of the hypothesized SDT dynamic model. The average age was 46.33 (SD=7.4) years, and the average steps per day at baseline was 5,507 steps (SD=6,239). The hypothesized 5-input SDT-inspired ARX model for app engagement resulted in a 31.75 % weighted RMSEA (31.50 % on validation and 31.91 % on estimation), indicating that the model predicted app page views almost 32 % better relative to the mean of the data. Among Hispanic/Latino participants, the average overall model fit across inventories of the SDT fluid analogy was 34.22 % (SD=10.53) compared to 22.39 % (SD=6.36) among non-Hispanic/Latino Whites, a difference of 11.83 %. Across individuals, the number of daily notification prompts received by the participant was positively associated with increased app page views. The weekend/weekday indicator and perceived daily busyness were also found to be key predictors of the number of daily application page views. CONCLUSIONS: This novel approach has significant implications for both personalized and adaptive DBCIs by identifying factors that foster or undermine engagement in an individual's respective context. Once identified, these factors can be tailored to promote engagement and support sustained behavior change over time.

Just-in-Time Adaptive Interventions (JITAIs) are a class of personalized health interventions developed within the behavioral science community. JITAIs aim to provide the right type and amount of support by iteratively selecting a sequence of intervention options from a pre-defined set of components in response to each individual’s time varying state. In this work, we explore the application of reinforcement learning methods to the problem of learning intervention option selection policies. We study the effect of context inference error and partial observability on the ability to learn effective policies. Our results show that the propagation of uncertainty from context inferences is critical to improving intervention efficacy as context uncertainty increases, while policy gradient algorithms can provide remarkable robustness to partially observed behavioral state information.

Visual object tracking has seen significant progress in recent years. However, the vast majority of this work focuses on tracking objects within the image plane of a single camera and ignores the uncertainty associated with predicted object locations. In this work, we focus on the geospatial object tracking problem using data from a distributed camera network. The goal is to predict an object’s track in geospatial coordinates along with uncertainty over the object’s location while respecting communication constraints that prohibit centralizing raw image data. We present a novel single-object geospatial tracking data set that includes high-accuracy ground truth object locations and video data from a network of four cameras. We present a modeling framework for addressing this task including a novel backbone model and explore how uncertainty calibration and fine-tuning through a differentiable tracker affect performance.

Sensing and data collection platforms are the crucial components of high-quality datasets that can fuel advancements in research. However, such platforms usually are ad-hoc designs and are limited in sensor modalities. In this paper, we discuss our experience designing and deploying a multi-modal multi-node sensor data collection platform that can be utilized for various data collection tasks. The main goal of this platform is to create a modality-rich data collection platform suitable for Internet of Things (IoT) applications with easy reproducibility and deployment, which can accelerate data collection and downstream research tasks.

Bayesian decision theory provides an elegant framework for acting optimally under uncertainty when tractable posterior distributions are available. Modern Bayesian models, however, typically involve intractable posteriors that are approximated with, potentially crude, surrogates. This difficulty has engendered loss-calibrated techniques that aim to learn posterior approximations that favor high-utility decisions. In this paper, focusing on Bayesian neural networks, we develop methods for correcting approximate posterior predictive distributions encouraging them to prefer high-utility decisions. In contrast to previous work, our approach is agnostic to the choice of the approximate inference algorithm, allows for efficient test time decision making through amortization, and empirically produces higher quality decisions. We demonstrate the effectiveness of our approach through controlled experiments spanning a diversity of tasks and datasets.

Irregularly-sampled time series occur in many domains including healthcare. They can be challenging to model because they do not naturally yield a fixed-dimensional representation as required by many standard machine learning models. In this paper, we consider irregular sampling from the perspective of missing data. We model observed irregularly-sampled time series data as a sequence of index-value pairs sampled from a continuous but unobserved function. We introduce an encoder-decoder framework for learning from such generic indexed sequences. We propose learning methods for this framework based on variational autoencoders and generative adversarial networks. For continuous irregularly-sampled time series, we introduce continuous convolutional layers that can efficiently interface with existing neural network architectures. Experiments show that our models are able to achieve competitive or better classification results on irregularly-sampled multivariate time series compared to recent RNN models while offering significantly faster training times.

Recent years have seen dramatic advances in low-power neural accelerators that aim to bring deep learning analytics to IoT devices; simultaneously, there have been considerable advances in the design of low-power radios to enable efficient compute offload from IoT devices to the cloud. Neither is a panacea --- deep learning models are often too large for low-power accelerators and bandwidth needs are often too high for low-power radios. While there has been considerable work on deep learning for smartphone-class devices, these methods do not work well for small battery-powered IoT devices that are considerably more resource-constrained.

In this paper, we present a general framework for distilling expectations with respect to the Bayesian posterior distribution of a deep neural network classifier, extending prior work on the Bayesian Dark Knowledge framework. The proposed framework takes as input "teacher" and "student" model architectures and a general posterior expectation of interest. The distillation method performs an online compression of the selected posterior expectation using iteratively generated Monte Carlo samples. We focus on the posterior predictive distribution and expected entropy as distillation targets. We investigate several aspects of this framework including the impact of uncertainty and the choice of student model architecture. We study methods for student model architecture search from a speed-storage-accuracy perspective and evaluate down-stream tasks leveraging entropy distillation including uncertainty ranking and out-of-distribution detection.

In mobile health (mHealth) and human activity recognition (HAR), collecting labeled data often comes at a significantly higher cost or level of user burden than collecting unlabeled data. This motivates the idea of attempting to optimize the collection of labeled data to minimize cost or burden. In this paper, we develop active learning methods that are tailored to the mHealth and HAR domains to address the problems of labeled data scarcity and the cost of labeled data collection. Specifically, we leverage between-user similarity to propose a novel hierarchical active learning framework that personalizes models for each user while sharing the labeled data collection burden across a group, thereby reducing the labeling effort required by any individual user. We evaluate our framework on a publicly available human activity recognition dataset. Our hierarchical active learning framework on average achieves between a 20% and 70% reduction in labeling effort when compared to standard active learning methods.

The ability to monitor eye closures and blink patterns has long been known to enable accurate assessment of fatigue and drowsiness in individuals. Many measures of the eye are known to be correlated with fatigue including coarse-grained measures like the rate of blinks as well as fine-grained measures like the duration of blinks and the extent of eye closures. Despite a plethora of research validating these measures, we lack wearable devices that can continually and reliably monitor them in the natural environment. In this work, we present a low-power system, iLid, that can continually sense fine-grained measures such as blink duration and Percentage of Eye Closures (PERCLOS) at high frame rates of 100fps. We present a complete solution including design of the sensing, signal processing, and machine learning pipeline and implementation on a prototype computational eyeglass platform.

Heating, ventilation, and air conditioning (HVAC) accounts for over 50% of a typical home’s energy usage. A thermostat generally controls HVAC usage in a home to ensure user comfort. In this article, we focus on making existing “dumb” programmable thermostats smart by applying energy analytics on smart meter data to infer home occupancy patterns and compute an optimized thermostat schedule. Utilities with smart meter deployments are capable of immediately applying our approach, called iProgram, to homes across their customer base. iProgram addresses new challenges in inferring home occupancy from smart meter data where (i) training data is not available and (ii) the thermostat schedule may be misaligned with occupancy, frequently resulting in high power usage during unoccupied periods. iProgram translates occupancy patterns inferred from opaque smart meter data into a custom schedule for existing types of programmable thermostats, e.g., 1-day, 7-day, and so on. We implement iProgram as a web service and show that it reduces the mismatch time between the occupancy pattern and the thermostat schedule by a median value of 44.28min (out of 100 homes) when compared to a default 8am-6pm weekday schedule, with a median deviation of 30.76min off the optimal schedule. Further, iProgram yields a daily energy savings of 0.42kWh on average across the 100 homes. Moreover, the schedules generated from iProgram converge to optimal schedules within a couple of weeks for most homes. We also show that homeowners having multiple HVAC zones can utilize iProgram and potentially increase unconditioned times of less occupied parts of their homes by 70%. Utilities may use iProgram to recommend thermostat schedules to customers and provide them estimates of potential energy savings in their energy bills.

We consider the ultimate limits of program-specific garbage collector (GC) performance for real programs. We first characterize the GC schedule optimization problem. Based on this characterization, we develop a linear-time dynamic programming solution that, given a program run and heap size, computes an optimal schedule of collections for a non-generational collector. Using an analysis of a heap object graph of the program, we compute a property of heap objects that we call their pre-birth time. This information enables us to extend the non-generational GC schedule problem to the generational GC case in a way that also admits a dynamic programming solution with cost quadratic in the length of the trace (number of objects allocated). This improves our previously reported approximately optimal result. We further extend the two-generation dynamic program to any number of generations, allowing other generalizations as well. Our experimental results for two generations on traces from Java programs of the DaCapo benchmark suite show that there is considerable promise to reduce garbage collection costs for some programs by developing program-specific collection policies. For a given space budget, optimal schedules often obtain modest but useful time savings, and for a given time budget, optimal schedules can obtain considerable space savings.

An upcoming frontier for distributed computing might literally save lives in future military operations. In civilian scenarios, significant efficiencies were gained from interconnecting devices into networked services and applications that automate much of everyday life from smart homes to intelligent transportation. The ecosystem of such applications and services is collectively called the Internet of Things (IoT). Can similar benefits be gained in a military context by developing an IoT for the battlefield? This paper describes unique challenges in such a context as well as potential risks, mitigation strategies, and benefits.

In this paper, we consider a new low-quality label learning problem: learning time series detection models from temporally imprecise labels. In this problem, the data consist of a set of input time series, and supervision is provided by a sequence of noisy time stamps corresponding to the occurrence of positive class events. Such temporally imprecise labels commonly occur in areas like mobile health research where human annotators are tasked with labeling the occurrence of very short duration events. We propose a general learning framework for this problem that can accommodate different base classifiers and noise models. We present results on real mobile health data showing that the proposed framework significantly outperforms a number of alternatives including assuming that the label time stamps are noise-free, transforming the problem into the multiple instance learning framework, and learning on labels that were manually re-aligned.

In this paper, we present a new approach to learning cascaded classifiers for use in computing environments that involve networks of heterogeneous and resource-constrained, low-power embedded compute and sensing nodes. We present a generalization of the classical linear detection cascade to the case of tree-structured cascades where different branches of the tree execute on different physical compute nodes in the network. Different nodes have access to different features, as well as access to potentially different computation and energy resources. We concentrate on the problem of jointly learning the parameters for all of the classifiers in the cascade given a fixed cascade architecture and a known set of costs required to carry out the computation at each node. To accomplish the objective of joint learning of all detectors, we propose a novel approach to combining classifier outputs during training that better matches the hard cascade setting in which the learned system will be deployed. This work is motivated by research in the area of mobile health where energy efficient real time detectors integrating information from multiple wireless on-body sensors and a smart phone are needed for real-time monitoring and the delivery of just-in-time adaptive interventions. We evaluate our framework on mobile sensor-based human activity recognition and mobile health detector learning problems.

This paper proposes an out-of-sample extension framework for a global manifold learning algorithm (Isomap) that uses temporal information in out-of-sample points in order to make the embedding more robust to noise and artifacts. Given a set of noise-free training data and its embedding, the proposed framework extends the embedding for a noisy time series. This is achieved by adding a spatio-temporal compactness term to the optimization objective of the embedding. To the best of our knowledge, this is the first method for out-of-sample extension of manifold embeddings that leverages timing information available for the extension set. Experimental results demonstrate that our out-of-sample extension algorithm renders a more robust and accurate embedding of sequentially ordered image data in the presence of various noise and artifacts when compared with other timing-aware embeddings. Additionally, we show that an out-of-sample extension framework based on the proposed algorithm outperforms the state of the art in eye-gaze estimation.

The human eye offers a fascinating window into an individual's health, cognitive attention, and decision making, but we lack the ability to continually measure these parameters in the natural environment. We demonstrate CIDER, a system that operates in a highly optimized low-power mode under indoor settings by using a fast Search-Refine controller to track the eye, but detects when the environment switches to more challenging outdoor sunlight and switches models to operate robustly under this condition. Our design is holistic and tackles a) power consumption in digitizing pixels, estimating pupillary parameters, and illuminating the eye via near-infrared and b) error in estimating pupil center and pupil dilation. We demonstrate that CIDER can estimate pupil center with error less than two pixels (0.6°), and pupil diameter with error of one pixel (0.22mm). Our end-to-end results show that we can operate at power levels of roughly 7mW at a 4Hz eye tracking rate, or roughly 32mW at rates upwards of 250Hz.

Mobile health research on illicit drug use detection typically involves a two-stage study design where data to learn detectors is first collected in lab-based trials, followed by a deployment to subjects in a free-living environment to assess detector performance. While recent work has demonstrated the feasibility of wearable sensors for illicit drug use detection in the lab setting, several key problems can limit lab-to-field generalization performance. For example, lab-based data collection often has low ecological validity, the ground-truth event labels collected in the lab may not be available at the same level of temporal granularity in the field, and there can be significant variability between subjects. In this paper, we present domain adaptation methods for assessing and mitigating potential sources of performance loss in lab-to-field generalization and apply them to the problem of cocaine use detection from wearable electrocardiogram sensor data.

The field of mobile health (mHealth) has the potential to yield new insights into health and behavior through the analysis of continuously recorded data from wearable health and activity sensors. In this paper, we present a hierarchical span-based conditional random field model for the key problem of jointly detecting discrete events in such sensor data streams and segmenting these events into high-level activity sessions. Our model includes higher-order cardinality factors and inter-event duration factors to capture domain-specific structure in the label space. We show that our model supports exact MAP inference in quadratic time via dynamic programming, which we leverage to perform learning in the structured support vector machine framework. We apply the model to the problems of smoking and eating detection using four real data sets. Our results show statistically significant improvements in segmentation performance relative to a hierarchical pairwise CRF.

We present a general framework for classification of sparse and irregularly-sampled time series. The properties of such time series can result in substantial uncertainty about the values of the underlying temporal processes, while making the data difficult to deal with using standard classification methods that assume fixed-dimensional feature spaces. To address these challenges, we propose an uncertainty-aware classification framework based on a special computational layer we refer to as the Gaussian process adapter that can connect irregularly sampled time series data to any black-box classifier learnable using gradient descent. We show how to scale up the required computations based on combining the structured kernel interpolation framework and the Lanczos approximation method, and how to discriminatively train the Gaussian process adapter in combination with a number of classifiers end-to-end using backpropagation.

We consider the ultimate limits of program-specific garbage collector performance for real programs. We first characterize the GC schedule optimization problem using Markov Decision Processes (MDPs). Based on this characterization, we develop a method of determining, for a given program run and heap size, an optimal schedule of collections for a non-generational collector. We further explore the limits of performance of a generational collector, where it is not feasible to search the space of schedules to prove optimality. Still, we show significant improvements with Least Squares Policy Iteration, a reinforcement learning technique for solving MDPs. We demonstrate that there is considerable promise to reduce garbage collection costs by developing program-specific collection policies.

Recent researches have demonstrated the feasibility of detecting smoking from wearable sensors, but their performance on real-life smoking lapse detection is unknown. In this paper, we propose a new model and evaluate its performance on 61 newly abstinent smokers for detecting a first lapse. We use two wearable sensors - breathing pattern from respiration and arm movements from 6-axis inertial sensors worn on wrists. In 10-fold cross-validation on 40 hours of training data from 6 daily smokers, our model achieves a recall rate of 96.9%, for a false positive rate of 1.1%. When our model is applied to 3 days of post-quit data from 32 lapsers, it correctly pinpoints the timing of first lapse in 28 participants. Only 2 false episodes are detected on 20 abstinent days of these participants. When tested on 84 abstinent days from 28 abstainers, the false episode per day is limited to 1/6.

The human eye offers a fascinating window into an individual's health, cognitive attention, and decision making, but we lack the ability to continually measure these parameters in the natural environment. The challenges lie in: a) handling the complexity of continuous high-rate sensing from a camera and processing the image stream to estimate eye parameters, and b) dealing with the wide variability in illumination conditions in the natural environment. This paper explores the power-robustness tradeoffs inherent in the design of a wearable eye tracker, and proposes a novel staged architecture that enables graceful adaptation across the spectrum of real-world illumination. We propose CIDER, a system that operates in a highly optimized low-power mode under indoor settings by using a fast Search-Refine controller to track the eye, but detects when the environment switches to more challenging outdoor sunlight and switches models to operate robustly under this condition. Our design is holistic and tackles a) power consumption in digitizing pixels, estimating pupillary parameters, and illuminating the eye via near-infrared, b) error in estimating pupil center and pupil dilation, and c) model training procedures that involve zero effort from a user. We demonstrate that CIDER can estimate pupil center with error less than two pixels (0.6°), and pupil diameter with error of one pixel (0.22mm). Our end-to-end results show that we can operate at power levels of roughly 7mW at a 4Hz eye tracking rate, or roughly 32mW at rates upwards of 250Hz.

Abstract We present a method for joint analysis and synthesis of geometrically diverse 3D shape families. Our method first learns part‐based templates such that an optimal set of fuzzy point and part correspondences is computed between the shapes of an input collection based on a probabilistic deformation model. In contrast to previous template‐based approaches, the geometry and deformation parameters of our part‐based templates are learned from scratch. Based on the estimated shape correspondence, our method also learns a probabilistic generative model that hierarchically captures statistical relationships of corresponding surface point positions and parts as well as their existence in the input shapes. A deep learning procedure is used to capture these hierarchical relationships. The resulting generative model is used to produce control point arrangements that drive shape synthesis by combining and deforming parts from the input collection. The generative model also yields compact shape descriptors that are used to perform fine‐grained classification. Finally, it can be also coupled with the probabilistic deformation model to further improve shape correspondence. We provide qualitative and quantitative evaluations of our method for shape correspondence, segmentation, fine‐grained classification and synthesis. Our experiments demonstrate superior correspondence and segmentation results than previous state‐of‐the‐art approaches.

Mobile sensor data-to-knowledge (MD2K) was chosen as one of 11 Big Data Centers of Excellence by the National Institutes of Health, as part of its Big Data-to-Knowledge initiative. MD2K is developing innovative tools to streamline the collection, integration, management, visualization, analysis, and interpretation of health data generated by mobile and wearable sensors. The goal of the big data solutions being developed by MD2K is to reliably quantify physical, biological, behavioral, social, and environmental factors that contribute to health and disease risk. The research conducted by MD2K is targeted at improving health through early detection of adverse health events and by facilitating prevention. MD2K will make its tools, software, and training materials widely available and will also organize workshops and seminars to encourage their use by researchers and clinicians.

We propose a novel discriminative model for semantic labeling in videos by incorporating a prior to model both the shape and temporal dependencies of an object in video. A typical approach for this task is the conditional random field (CRF), which can model local interactions among adjacent regions in a video frame. Recent work has shown how to incorporate a shape prior into a CRF for improving labeling performance, but it may be difficult to model temporal dependencies present in video by using this prior. The conditional restricted Boltzmann machine (CRBM) can model both shape and temporal dependencies, and has been used to learn walking styles from motion- capture data. In this work, we incorporate a CRBM prior into a CRF framework and present a new state-of-the-art model for the task of semantic labeling in videos. In particular, we explore the task of labeling parts of complex face scenes from videos in the YouTube Faces Database (YFDB). Our combined model outperforms competitive baselines both qualitatively and quantitatively.

Continuous, real-time tracking of eye gaze is valuable in a variety of scenarios including hands-free interaction with the physical world, detection of unsafe behaviors, leveraging visual context for advertising, life logging, and others. While eye tracking is commonly used in clinical trials and user studies, it has not bridged the gap to everyday consumer use. The challenge is that a real-time eye tracker is a power-hungry and computation-intensive device which requires continuous sensing of the eye using an imager running at many tens of frames per second, and continuous processing of the image stream using sophisticated gaze estimation algorithms. Our key contribution is the design of an eye tracker that dramatically reduces the sensing and computation needs for eye tracking, thereby achieving orders of magnitude reductions in power consumption and form-factor. The key idea is that eye images are extremely redundant, therefore we can estimate gaze by using a small subset of carefully chosen pixels per frame. We use a sparse pixel-based gaze estimation algorithm that is a multi-layer neural network learned using a state-of-the-art sparsity-inducing regularization function which minimizes the gaze prediction error while simultaneously minimizing the number of pixels used. Our results show that we can operate at roughly 70mW of power, while continuously estimating eye gaze at the rate of 30 Hz with errors of roughly 4 degrees.

Continuous, real-time tracking of eye gaze is valuable in a variety of scenarios including hands-free interaction with the physical world, detection of unsafe behaviors, leveraging visual context for advertising, life logging, and others. While eye tracking is commonly used in clinical trials and user studies, it has not bridged the gap to everyday consumer use. The challenge is that a real-time eye tracker is a power-hungry and computation-intensive device which requires continuous sensing of the eye using an imager running at many tens of frames per second, and continuous processing of the image stream using sophisticated gaze estimation algorithms. Our key contribution is the design of an eye tracker that dramatically reduces the sensing and computation needs for eye tracking, thereby achieving orders of magnitude reductions in power consumption and form-factor. The key idea is that eye images are extremely redundant, therefore we can estimate gaze by using a small subset of carefully chosen pixels per frame. We instantiate this idea in a prototype hardware platform equipped with a low-power image sensor that provides random access to pixel values, a low-power ARM Cortex M3 microcontroller, and a bluetooth radio to communicate with a mobile phone. The sparse pixel-based gaze estimation algorithm is a multi-layer neural network learned using a state-of-the-art sparsity-inducing regularization function that minimizes the gaze prediction error while simultaneously minimizing the number of pixels used. Our results show that we can operate at roughly 70mW of power, while continuously estimating eye gaze at the rate of 30 Hz with errors of roughly 3 degrees.

The goal of computer tailored health communications (CTHC) is to elicit healthy behavior changes by sending motivational messages personalized to individual patients. One prominent weakness of many existing CTHC systems is that they are based on expert-written rules and thus have no ability to learn from their users over time. One solution to this problem is to develop CTHC systems based on the principles of collaborative filtering, but this approach has not been widely studied. In this paper, we present a case study evaluating nine rating prediction methods for use in the Patient Experience Recommender System for Persuasive Communication Tailoring, a system developed for use in a clinical trial of CTHC-based smoking cessation support interventions.

Ubiquitous physiological sensing has the potential to profoundly improve our understanding of human behavior, leading to more targeted treatments for a variety of disorders. The long term goal of this work is development of novel computational tools to support the study of addiction in the context of cocaine use. The current paper takes the first step in this important direction by posing a simple, but crucial question: Can cocaine use be reliably detected using wearable electrocardiogram (ECG) sensors? The main contributions in this paper include the presentation of a novel clinical study of cocaine use, the development of a computational pipeline for inferring morphological features from noisy ECG waveforms, and the evaluation of feature sets for cocaine use detection. Our results show that 32mg/70kg doses of cocaine can be detected with the area under the receiver operating characteristic curve levels above 0.9 both within and between-subjects.

Mobile phones have evolved from communication devices to indispensable accessories with access to real-time content. The increasing reliance on dynamic content comes at the cost of increased latency to pull the content from the Internet before the user can start using it. While prior work has explored parts of this problem, they ignore the bandwidth costs of prefetching, incur significant training overhead, need several sensors to be turned on, and do not consider practical systems issues that arise from the limited background processing capability supported by mobile operating systems. In this paper, we make app prefetch practical on mobile phones. Our contributions are two-fold. First, we design an app prediction algorithm, APPM, that requires no prior training, adapts to usage dynamics, predicts not only which app will be used next but also when it will be used, and provides high accuracy without requiring additional sensor context. Second, we perform parallel prefetch on screen unlock, a mechanism that leverages the benefits of prediction while operating within the constraints of mobile operating systems. Our experiments are conducted on long-term traces, live deployments on the Android Play Market, and user studies, and show that we outperform prior approaches to predicting app usage, while also providing practical ways to prefetch application content on mobile phones.

The proliferation of sensors on mobile phones and wearables has led to a plethora of context classifiers designed to sense the individual's context. We argue that a key missing piece in mobile inference is a layer that fuses the outputs of several classifiers to learn deeper insights into an individual's habitual patterns and associated correlations between contexts, thereby enabling new systems optimizations and opportunities. In this paper, we design CQue, a dynamic bayesian network that operates over classifiers for individual contexts, observes relations across these outputs across time, and identifies opportunities for improving energy-efficiency and accuracy by taking advantage of relations. In addition, such a layer provides insights into privacy leakage that might occur when seemingly innocuous user context revealed to different applications on a phone may be combined to reveal more information than originally intended. In terms of system architecture, our key contribution is a clean separation between the detection layer and the fusion layer, enabling classifiers to solely focus on detecting the context, and leverage temporal smoothing and fusion mechanisms to further boost performance by just connecting to our higher-level inference engine. To applications and users, CQue provides a query interface, allowing a) applications to obtain more accurate context results while remaining agnostic of what classifiers/sensors are used and when, and b) users to specify what contexts they wish to keep private, and only allow information that has low leakage with the private context to be revealed. We implemented CQue in Android, and our results show that CQue can i) improve activity classification accuracy up to 42%, ii) reduce energy consumption in classifying social, location and activity contexts with high accuracy(>90%) by reducing the number of required classifiers by at least 33%, and iii) effectively detect and suppress contexts that reveal private information.

The goal of rating-based recommender systems is to make personalized predictions and recommendations for individual users by leveraging the preferences of a community of users with respect to a collection of items like songs or movies. Recommender systems are often based on intricate statistical models that are estimated from data sets containing a very high proportion of missing ratings. This work describes evidence of a basic incompatibility between the properties of recommender system data sets and the assumptions required for valid estimation and evaluation of statistical models in the presence of missing data. We discuss the implications of this problem and describe extended modelling and evaluation frameworks that attempt to circumvent it. We present prediction and ranking results showing that models developed and tested under these extended frameworks can significantly outperform standard models. 1

We propose a new variational EM algorithm for fitting factor analysis models with mixed continuous and categorical observations. The algorithm is based on a simple quadratic bound to the log-sum-exp function. In the special case of fully observed binary data, the bound we propose is significantly faster than previous variational methods. We show that EM is significantly more robust in the presence of missing data compared to treating the latent factors as parameters, which is the approach used by exponential family PCA and other related matrix-factorization methods. A further benefit of the variational approach is that it can easily be extended to the case of mixtures of factor analyzers, as we show. We present results on synthetic and real data sets demonstrating several desirable properties of our proposed method.

Recent work has shown that one can learn the structure of Gaussian Graphical Models by imposing an L1 penalty on the precision matrix, and then using efficient convex optimization methods to find the penalized maximum likelihood estimate. This is similar to performing MAP estimation with a prior that prefers sparse graphs. In this paper, we use the stochastic block model as a prior. This prefer graphs that are blockwise sparse, but unlike previous work, it does not require that the blocks or groups be specified a priori. The resulting problem is no longer convex, but we devise an efficient variational Bayes algorithm to solve it. We show that our method has better test set likelihood on two different datasets (motion capture and gene expression) compared to independent L1, and can match the performance of group L1 using manually created groups.

In this paper we make several contributions towards accelerating approximate Bayesian structural inference for non-decomposable GGMs. Our first contribution is to show how to efficiently compute a BIC or Laplace approximation to the marginal likelihood of non-decomposable graphs using convex methods for precision matrix estimation. This optimization technique can be used as a fast scoring function inside standard Stochastic Local Search (SLS) for generating posterior samples. Our second contribution is a novel framework for efficiently generating large sets of high-quality graph topologies without performing local search. This graph proposal method, which we call Neighborhood Fusion" (NF), samples candidate Markov blankets at each node using sparse regression techniques. Our final contribution is a hybrid method combining the complementary strengths of NF and SLS. Experimental results in structural recovery and prediction tasks demonstrate that NF and hybrid NF/SLS out-perform state-of-the-art local search methods, on both synthetic and real-world datasets, when realistic computational limits are imposed."

A fundamental aspect of rating-based recommender systems is the observation process, the process by which users choose the items they rate. Nearly all research on collaborative filtering and recommender systems is founded on the assumption that missing ratings are missing at random. The statistical theory of missing data shows that incorrect assumptions about missing data can lead to biased parameter estimation and prediction. In a recent study, we demonstrated strong evidence for violations of the missing at random condition in a real recommender system. In this paper we present the first study of the effect of non-random missing data on collaborative ranking, and extend our previous results regarding the impact of non-random missing data on collaborative prediction.

We describe a class of causal, discrete latent variable models called Multiple Multiplicative Factor models (MMFs). A data vector is represented in the latent space as a vector of factors that have discrete, non-negative expression levels. Each factor proposes a distribution over the data vector. The distinguishing feature of MMFs is that they combine the factors' proposed distributions multiplicatively, taking into account factor expression levels. The product formulation of MMFs allow factors to specialize to a subset of the items, while the causal generative semantics mean MMFs can readily accommodate missing data. This makes MMFs distinct from both directed models with mixture semantics and undirected product models. In this paper we present empirical results from the collaborative filtering domain showing that a binary/multinomial MMF model matches the performance of the best existing models while learning an interesting latent space description of the users.

In this paper we present a generative latent variable model for rating-based collaborative (cid:12)ltering called the User Rating Pro(cid:12)le model (URP). The generative process which underlies URP is de- signed to produce complete user rating pro(cid:12)les, an assignment of one rating to each item for each user. Our model represents each user as a mixture of user attitudes, and the mixing proportions are distributed according to a Dirichlet random variable. The rating for each item is generated by selecting a user attitude for the item, and then selecting a rating according to the preference pattern associ- ated with that attitude. URP is related to several models including a multinomial mixture model, the aspect model [7], and LDA [1], but has clear advantages over each.