This work is motivated by the engineering task of achieving a near state-of-the-art face recognition on a minimal computing budget running on an embedded system. Our main technical contribution centers around a novel training method, called Multibatch, for similarity learning, i.e., for the task of generating an invariant "face signature" through training pairs of "same" and "not-same" face images. The Multibatch method first generates signatures for a mini-batch of k face images and then constructs an unbiased estimate of the full gradient by relying on all k 2 − k pairs from the mini-batch. We prove that the variance of the Multibatch estimator is bounded by O(1/k 2 ), under some mild conditions. In contrast, the standard gradient estimator that relies on random k/2 pairs has a variance of order 1/k. The smaller variance of the Multibatch estimator significantly speeds up the convergence rate of stochastic gradient descent. Using the Multibatch method we train a deep convolutional neural network that achieves an accuracy of 98.2% on the LFW benchmark, while its prediction runtime takes only 30msec on a single ARM Cortex A9 core. Furthermore, the entire training process took only 12 hours on a single Titan X GPU.

Sleep is associated with important changes in respiratory rate and ventilation. Currently, breathing rate (BR) is measured during sleep using an array of contact and wearable sensors, including airflow sensors and respiratory belts; there is need for a simplified and more comfortable approach to monitor respiration. Here, we present a new method for BR evaluation during sleep using a non-contact microphone. The basic idea behind this approach is that during sleep the upper airway becomes narrower due to muscle relaxation, which leads to louder breathing sounds that can be captured via ambient microphone. In this study we developed a signal processing algorithm that emphasizes breathing sounds, extracts breathing-related features, and estimates BR during sleep. A comparison between audio-based BR estimation and BR calculated using the traditional (gold-standard) respiratory belts during in-laboratory polysomnography (PSG) study was performed on 204 subjects. Pearson's correlation between subjects' averaged BR of the two approaches was R=0.97. Epoch-by-epoch (30 s) BR comparison revealed a mean relative error of 2.44% and Pearson's correlation of 0.68. This study shows reliable and promising results for non-contact BR estimation.

Evaluation of respiratory activity during sleep is essential in order to reliably diagnose sleep disorder breathing (SDB); a condition associated with serious cardio-vascular morbidity and mortality. In the current study, we developed and validated a robust automatic breathing-sounds (i.e. inspiratory and expiratory sounds) detection system of audio signals acquired during sleep. Random forest classifier was trained and tested using inspiratory/expiratory/noise events (episodes), acquired from 84 subjects consecutively and prospectively referred to SDB diagnosis in sleep laboratory and in at-home environment. More than 560,000 events were analyzed, including a variety of recording devices and different environments. The system's overall accuracy rate is 88.8%, with accuracy rate of 91.2% and 83.6% in in-laboratory and at-home environments respectively, when classifying between inspiratory, expiratory, and noise classes. Here, we provide evidence that breathing-sounds can be reliably detected using non-contact audio technology in at-home environment. The proposed approach may improve our understanding of respiratory activity during sleep. This in return, will improve early SDB diagnosis and treatment.

Obstructive sleep apnea (OSA) is a prevalent sleep disorder, characterized by recurrent episodes of upper airway obstructions during sleep. We hypothesize that breath-by-breath audio analysis of the respiratory cycle (i.e., inspiration and expiration phases) during sleep can reliably estimate the apnea hypopnea index (AHI), a measure of OSA severity. The AHI is calculated as the average number of apnea (A)/hypopnea (H) events per hour of sleep. Audio signals recordings of 186 adults referred to OSA diagnosis were acquired in-laboratory and at-home conditions during polysomnography and WatchPat study, respectively. A/H events were automatically segmented and classified using a binary random forest classifier. Total accuracy rate of 86.3% and an agreement of κ=42.98% were achieved in A/H event detection. Correlation of r=0.87 (r=0.74), diagnostic agreement of 76% (81.7%), and average absolute difference AHI error of 7.4 (7.8) (events/hour) were achieved in in-laboratory (at-home) conditions, respectively. Here we provide evidence that A/H events can be reliably detected at their exact time locations during sleep using non-contact audio approach. This study highlights the potential of this approach to reliably evaluate AHI in at home conditions.