Inter-beat interval (IBI) measurement enables estimation of heart-tare variability (HRV) which, in turn, can provide early indication of potential cardiovascular diseases (CVDs). However, extracting IBIs from noisy signals is challenging since the morphology of the signal gets distorted in the presence of noise. Electrocardiogram (ECG) of a person in heavy motion is highly corrupted with noise, known as motion-artifact, and IBI extracted from it is inaccurate. As a part of remote health monitoring and wearable system development, denoising ECG signals and estimating IBIs correctly from them have become an emerging topic among signal-processing researchers. Apart from conventional methods, deep-learning techniques have been successfully used in signal denoising recently, and diagnosis process has become easier, leading to accuracy levels that were previously unachievable. We propose a deep-learning approach leveraging tiramisu autoencoder model to suppress motion-artifact noise and make the R-peaks of the ECG signal prominent even in the presence of high-intensity motion. After denoising, IBIs are estimated more accurately expediting diagnosis tasks. Results illustrate that our method enables IBI estimation from noisy ECG signals with SNR up to -30 dB with average root mean square error (RMSE) of 13 milliseconds for estimated IBIs. At this noise level, our error percentage remains below 8% and outperforms other state-of-the-art techniques.

A primary focus area in continual learning research is alleviating the "catastrophic forgetting" problem in neural networks by designing new algorithms that are more robust to the distribution shifts. While the recent progress in continual learning literature is encouraging, our understanding of what properties of neural networks contribute to catastrophic forgetting is still limited. To address this, instead of focusing on continual learning algorithms, in this work, we focus on the model itself and study the impact of "width" of the neural network architecture on catastrophic forgetting, and show that width has a surprisingly significant effect on forgetting. To explain this effect, we study the learning dynamics of the network from various perspectives such as gradient orthogonality, sparsity, and lazy training regime. We provide potential explanations that are consistent with the empirical results across different architectures and continual learning benchmarks.

Wearables are poised to transform health and wellness through automation of cost-effective, objective, and real-time health monitoring. However, machine learning models for these systems are designed based on labeled data collected, and feature representations engineered, in controlled environments. This approach has limited scalability of wearables because (i) collecting and labeling sufficiently large amounts of sensor data is a labor-intensive and expensive process; and (ii) wearables are deployed in highly dynamic environments of the end-users whose context undergoes consistent changes. We introduce TransNet , a deep learning framework that minimizes the costly process of data labeling, feature engineering, and algorithm retraining by constructing a scalable computational approach. TransNet learns general and reusable features in lower layers of the framework and quickly reconfigures the underlying models from a small number of labeled instances in a new domain, such as when the system is adopted by a new user or when a previously unseen event is to be added to event vocabulary of the system. Utilizing TransNet on four activity datasets, TransNet achieves an average accuracy of 88.1% in cross-subject learning scenarios using only one labeled instance for each activity class. This performance improves to an accuracy of 92.7% with five labeled instances.

Despite the fact that deep neural networks are powerful models and achieve appealing results on many tasks, they are too large to be deployed on edge devices like smartphones or embedded sensor nodes. There have been efforts to compress these networks, and a popular method is knowledge distillation, where a large (teacher) pre-trained network is used to train a smaller (student) network. However, in this paper, we show that the student network performance degrades when the gap between student and teacher is large. Given a fixed student network, one cannot employ an arbitrarily large teacher, or in other words, a teacher can effectively transfer its knowledge to students up to a certain size, not smaller. To alleviate this shortcoming, we introduce multi-step knowledge distillation, which employs an intermediate-sized network (teacher assistant) to bridge the gap between the student and the teacher. Moreover, we study the effect of teacher assistant size and extend the framework to multi-step distillation. Theoretical analysis and extensive experiments on CIFAR-10,100 and ImageNet datasets and on CNN and ResNet architectures substantiate the effectiveness of our proposed approach.

With the recent advances in both machine learning and embedded systems research, the demand to deploy computational models for real-time execution on edge devices has increased substantially. Without deploying computational models on edge devices, the frequent transmission of sensor data to the cloud results in rapid battery draining due to the energy consumption of wireless data transmission. This rapid power dissipation leads to a considerable reduction in the battery lifetime of the system, therefore jeopardizing the real-world utility of smart devices. It is well-established that for difficult machine learning tasks, models with higher performance often require more computation power and thus are not power-efficient choices for deployment on edge devices. However, the trade-offs between performance and power consumption are not well studied. While numerous methods (e.g., model compression) have been developed to obtain an optimal model, these methods focus on improving the efficiency of a "single" model. In an entirely new direction, we introduce an effective method to find a combination of "multiple" models that are optimal in terms of power-efficiency and performance by solving an optimization problem in which both performance and power consumption are taken into account. Experimental results demonstrate that on the ImageNet dataset, we can achieve a 20% energy reduction with only 0.3% accuracy drop compared to Squeeze-and-Excitation Networks. Compared to a pruned convolutional neural network for human activity recognition, while consuming 1.7% less energy, our proposed policy achieves 1.3% higher accuracy.

Catastrophic forgetting affects the training of neural networks, limiting their ability to learn multiple tasks sequentially. From the perspective of the well established plasticity-stability dilemma, neural networks tend to be overly plastic, lacking the stability necessary to prevent the forgetting of previous knowledge, which means that as learning progresses, networks tend to forget previously seen tasks. This phenomenon coined in the continual learning literature, has attracted much attention lately, and several families of approaches have been proposed with different degrees of success. However, there has been limited prior work extensively analyzing the impact that different training regimes -- learning rate, batch size, regularization method-- can have on forgetting. In this work, we depart from the typical approach of altering the learning algorithm to improve stability. Instead, we hypothesize that the geometrical properties of the local minima found for each task play an important role in the overall degree of forgetting. In particular, we study the effect of dropout, learning rate decay, and batch size on forming training regimes that widen the tasks' local minima and consequently, on helping it not to forget catastrophically. Our study provides practical insights to improve stability via simple yet effective techniques that outperform alternative baselines.