Conversational systems enable numerous valuable applications, and question-answering is an important component underlying many of these.However, conversational questionanswering remains challenging due to the lack of realistic, domain-specific training data.Inspired by this bottleneck, we focus on conversational question generation as a means to generate synthetic conversations for training and evaluation purposes.We present a number of novel strategies to improve conversational flow and accommodate varying question types and overall fluidity.Specifically, we design ChainCQG as a two-stage architecture that learns question-answer representations across multiple dialogue turns using a flow propagation training strategy.ChainCQG significantly outperforms both answer-aware and answer-unaware SOTA baselines (e.g., up to 48% BLEU-1 improvement).Additionally, our model is able to generate different types of questions, with improved fluidity and coreference alignment.

This paper presents the Active Structural Occupant Detector, an active vibration sensing system that detects stationary occupants through injection of vibration signals into the floor. Many smart buildings require occupant detection in order to provide personalized services. Some examples include optimized energy usage and security. Several methods currently exist for occupant detection, each with their own drawbacks, such as installation requirements. Structural vibration sensing overcomes many of these drawbacks by measuring impulses created by occupants to infer their movements, but cannot detect stationary occupants. The ASOD utilizes active vibration sources, which inject acoustic waves into the structure then measure how the structure responds to them. Any occupants present interact with these waves, causing changes to the measured signal. By characterizing the changes in how the waves travel, we can predict the presence or lack of an occupant with up to a 97.7% accuracy, as demonstrated by experiments in a real- world environment.

Gait health monitoring is critical for condition diagnosis and fall prediction in elderly populations. Existing methods for gait health monitoring (e.g. direct observation and sensing) are not suitable for non-clinical environments due to qualitative assessments or operational limitations. Our method utilizes footstep-induced floor vibration sensing to provide a passive gait health monitoring platform that can be used in non-clinical environments (e.g. home settings) to provide gait health information in a timely manner. We decompose vibration responses to obtain signal peaks that correspond to temporal gait information and leverage foot dominance to learn a signal amplitude-footstep ground reaction force transfer function. Preliminary results show that temporal gait parameters can be estimated with up to 99% accuracy and gait balance symmetry can be estimated with as low as 10.4% error.

Deep learning models have been successfully used in medical image analysis problems but they require a large amount of labeled images to obtain good performance. However, such large labeled datasets are costly to acquire. Active learning techniques can be used to minimize the number of required training labels while maximizing the model's performance. In this work, we propose a novel sampling method that queries the unlabeled examples that maximize the average distance to all training set examples in a learned feature space. We then extend our sampling method to define a better initial training set, without the need for a trained model, by using Oriented FAST and Rotated BRIEF (ORB) feature descriptors. We validate MedAL on 3 medical image datasets and show that our method is robust to different dataset properties. MedAL is also efficient, achieving 80% accuracy on the task of Diabetic Retinopathy detection using only 425 labeled images, corresponding to a 32% reduction in the number of required labeled examples compared to the standard uncertainty sampling technique, and a 40% reduction compared to random sampling.

No abstract available.

Touch surfaces are intuitive interfaces for computing devices. Most of the traditional touch interfaces (vision, IR, capacitive, etc.) have mounting requirements, resulting in specialized touch surfaces limited by their size, cost, and mobility. More recent work has shown that vibration-based touch sensing techniques can localize taps/knocks, which provides a low-cost flexible alternative. These surfaces are envisioned as intuitive inputs for applications such as interactive meeting tables, smart kitchen appliance control, etc. However, due to dispersive and reflective properties of various vibrating mediums, it is difficult to localize taps accurately on ubiquitous surfaces. Furthermore, no work has been done on tracking continuous swipe interactions through vibration sensing.

We introduce a sensing system which leverages footstep-induced structural vibration for occupant localization. Such localization is important for many smart building applications, such as efficient building management, senior/health care, and security. Compared to other sensing approaches, footstep-induced vibration provides a sensing system which is sparse and non-intrusive. The main challenge of achieving high accuracy using such approach is frequency-dependent wave propagation characteristics, such as dispersion, in floor structure. These characteristics result in distortions in the shape of signal. To overcome such distortions, we decompose the vibration signals into different frequency components using wavelet transform and focus on specific components in all the sensors. In a set of experiments in a real structure, our approach results in average localization errors of 0.41 meters, a 4.4X reduction compared to a baseline approach using raw data.

Tracking multiple people in an indoor environment enables various smart building applications such as HVAC energy saving, patient/child monitoring, etc. Researchers have explored various sensing methods including vision, motion, and RF, which either require specific installation requirements or high deployment density. We introduce a passive sparse sensing method based on ambient structural vibration induced by foot strikes. Our system tracks multiple people based on the premise that human foot strikes have spatio-temporal variation, and hence do not fully overlap. The system achieved less than 0.4m accuracy in both one and two persons stepping conditions.

We describe a step-level indoor localization system which uses the ground vibration induced by human footsteps. Indoor localization is important for various smart building applications, including resources arrangement optimization, patient/customer tracking, etc. Geophones are used to measure the ground vibrations and time difference of arrival (TDoA) for different sensors are used to solve the multilateration localization problem. The advantages of this system include its sparsity and also its stability over time. Lesser dependency on instrument people is another upside of this system. The results of pilot tests show that this system can be successfully used for indoor localization.

The dynamic responses of the structure provide a variety of information about the structure as well as people inside. Compared with traditional sensing methods, structural sensing method serves more general sensing purposes due to the diversity of information it can infer from structural responses. For example, by sensing the structural vibration, a system can track and identify a person through vibration caused by their gaits [5, 6]. Such non-intrusive identification and tracking system enables various smart building applications, such as patient monitoring system at advanced hospitals and nursing homes.