Batteryless or so called passive wearables are providing new and innovative methods for human activity recognition (HAR), especially in healthcare applications for older people. Passive sensors are low cost, lightweight, unobtrusive and desirably disposable; attractive attributes for healthcare applications in hospitals and nursing homes. Despite the compelling propositions for sensing applications, the data streams from these sensors are characterised by high sparsity---the time intervals between sensor readings are irregular while the number of readings per unit time are often limited. In this paper, we rigorously explore the problem of learning activity recognition models from temporally sparse data. We describe how to learn directly from sparse data using a deep learning paradigm in an end-to-end manner. We demonstrate significant classification performance improvements on real-world passive sensor datasets from older people over the state-of-the-art deep learning human activity recognition models. Further, we provide insights into the model's behaviour through complementary experiments on a benchmark dataset and visualisation of the learned activity feature spaces.

We present practical Levenberg-Marquardt variants of Gauss-Newton and natural gradient methods for solving non-convex optimization problems that arise in training deep neural networks involving enormous numbers of variables and huge data sets. Our methods use subsampled Gauss-Newton or Fisher information matrices and either subsampled gradient estimates (fully stochastic) or full gradients (semi-stochastic), which, in the latter case, we prove convergent to a stationary point. By using the Sherman-Morrison-Woodbury formula with automatic differentiation (backpropagation) we show how our methods can be implemented to perform efficiently. Finally, numerical results are presented to demonstrate the effectiveness of our proposed methods.

Residual networks (ResNet) and weight normalization play an important role in various deep learning applications. However, parameter initialization strategies have not been studied previously for weight normalized networks and, in practice, initialization methods designed for un-normalized networks are used as a proxy. Similarly, initialization for ResNets have also been studied for un-normalized networks and often under simplified settings ignoring the shortcut connection. To address these issues, we propose a novel parameter initialization strategy that avoids explosion/vanishment of information across layers for weight normalized networks with and without residual connections. The proposed strategy is based on a theoretical analysis using mean field approximation. We run over 2,500 experiments and evaluate our proposal on image datasets showing that the proposed initialization outperforms existing initialization methods in terms of generalization performance, robustness to hyper-parameter values and variance between seeds, especially when networks get deeper in which case existing methods fail to even start training. Finally, we show that using our initialization in conjunction with learning rate warmup is able to reduce the gap between the performance of weight normalized and batch normalized networks.

Image reconstruction in low-count PET is particularly challenging because gammas from natural radioactivity in Lu-based crystals cause high random fractions that lower the measurement signal-to-noise-ratio (SNR). In model-based image reconstruction (MBIR), using more iterations of an unregularized method may increase the noise, so incorporating regularization into the image reconstruction is desirable to control the noise. New regularization methods based on learned convolutional operators are emerging in MBIR. We modify the architecture of a variational neural network, BCD-Net, for PET MBIR, and demonstrate the efficacy of the trained BCD-Net using XCAT phantom data that simulates the low true coincidence count-rates with high random fractions typical for Y-90 PET patient imaging after Y-90 microsphere radioembolization. Numerical results show that the proposed BCD-Net significantly improves PET reconstruction performance compared to MBIR methods using non-trained regularizers, total variation (TV) and non-local means (NLM), and a non-MBIR method using a single forward pass deep neural network, U-Net. BCD-Net improved activity recovery for a hot sphere significantly and reduced noise, whereas non-trained regularizers had a trade-off between noise and quantification. BCD-Net improved CNR and RMSE by 43.4% (85.7%) and 12.9% (29.1%) compared to TV (NLM) regularized MBIR. Moreover, whereas the image reconstruction results show that the non-MBIR U-Net over-fits the training data, BCD-Net successfully generalizes to data that differs from training data. Improvements were also demonstrated for the clinically relevant phantom measurement data where we used training and testing datasets having very different activity distribution and count-level.

Graph data widely exist in many high-impact applications. Inspired by the success of deep learning in grid-structured data, graph neural network models have been proposed to learn powerful node-level or graph-level representation. However, most of the existing graph neural networks suffer from the following limitations: (1) there is limited analysis regarding the graph convolution properties, such as seed-oriented, degree-aware and order-free; (2) the node's degree-specific graph structure is not explicitly expressed in graph convolution for distinguishing structure-aware node neighborhoods; (3) the theoretical explanation regarding the graph-level pooling schemes is unclear. To address these problems, we propose a generic degree-specific graph neural network named DEMO-Net motivated by Weisfeiler-Lehman graph isomorphism test that recursively identifies 1-hop neighborhood structures. In order to explicitly capture the graph topology integrated with node attributes, we argue that graph convolution should have three properties: seed-oriented, degree-aware, order-free. To this end, we propose multi-task graph convolution where each task represents node representation learning for nodes with a specific degree value, thus leading to preserving the degree-specific graph structure. In particular, we design two multi-task learning methods: degree-specific weight and hashing functions for graph convolution. In addition, we propose a novel graph-level pooling/readout scheme for learning graph representation provably lying in a degree-specific Hilbert kernel space. The experimental results on several node and graph classification benchmark data sets demonstrate the effectiveness and efficiency of our proposed DEMO-Net over state-of-the-art graph neural network models.

Recently, a parametrized class of loss functions called $\alpha$-loss, $\alpha \in [1,\infty]$, has been introduced for classification. This family, which includes the log-loss and the 0-1 loss as special cases, comes with compelling properties including an equivalent margin-based form which is classification-calibrated for all $\alpha$. We introduce a generalization of this family to the entire range of $\alpha \in (0,\infty]$ and establish how the parameter $\alpha$ enables the practitioner to choose among a host of operating conditions that are important in modern machine learning tasks. We prove that smaller $\alpha$ values are more conducive to faster optimization; in fact, $\alpha$-loss is convex for $\alpha \le 1$ and quasi-convex for $\alpha >1$. Moreover, we establish bounds to quantify the degradation of the local-quasi-convexity of the optimization landscape as $\alpha$ increases; we show that this directly translates to a computational slow down. On the other hand, our theoretical results also suggest that larger $\alpha$ values lead to better generalization performance. This is a consequence of the ability of the $\alpha$-loss to limit the effect of less likely data as $\alpha$ increases from 1, thereby facilitating robustness to outliers and noise in the training data. We provide strong evidence supporting this assertion with several experiments on benchmark datasets that establish the efficacy of $\alpha$-loss for $\alpha > 1$ in robustness to errors in the training data. Of equal interest is the fact that, for $\alpha < 1$, our experiments show that the decreased robustness seems to counteract class imbalances in training data.

Recent breakthroughs in defenses against adversarial examples, like adversarial training, make the neural networks robust against various classes of attackers (e.g., first-order gradient-based attacks). However, it is an open question whether the adversarially trained networks are truly robust under unknown attacks. In this paper, we present interval attacks, a new technique to find adversarial examples to evaluate the robustness of neural networks. Interval attacks leverage symbolic interval propagation, a bound propagation technique that can exploit a broader view around the current input to locate promising areas containing adversarial instances, which in turn can be searched with existing gradient-guided attacks. We can obtain such a broader view using sound bound propagation methods to track and over-approximate the errors of the network within given input ranges. Our results show that, on state-of-the-art adversarially trained networks, interval attack can find on average 47% relatively more violations than the state-of-the-art gradient-guided PGD attack.

Neural networks are easier to optimise when they have many more weights than are required for modelling the mapping from inputs to outputs. This suggests a two-stage learning procedure that first learns a large net and then prunes away connections or hidden units. But standard training does not necessarily encourage nets to be amenable to pruning. We introduce targeted dropout, a method for training a neural network so that it is robust to subsequent pruning. Before computing the gradients for each weight update, targeted dropout stochastically selects a set of units or weights to be dropped using a simple self-reinforcing sparsity criterion and then computes the gradients for the remaining weights. The resulting network is robust to post hoc pruning of weights or units that frequently occur in the dropped sets. The method improves upon more complicated sparsifying regularisers while being simple to implement and easy to tune.

This paper focuses on the discrimination capacity of aggregation functions: these are the permutation invariant functions used by graph neural networks to combine the features of nodes. Realizing that the most powerful aggregation functions suffer from a dimensionality curse, we consider a restricted setting. In particular, we show that the standard sum and a novel histogram-based function have the capacity to discriminate between any fixed number of inputs chosen by an adversary. Based on our insights, we design a graph neural network aiming, not to maximize discrimination capacity, but to learn discriminative graph representations that generalize well. Our empirical evaluation provides evidence that our choices can yield benefits to the problem of structural graph classification.

In this paper, we study how to solve resource allocation problems in ultra-reliable and low-latency communications by unsupervised deep learning, which often yield functional optimization problems with quality-of-service (QoS) constraints. We take a joint power and bandwidth allocation problem as an example, which minimizes the total bandwidth required to guarantee the QoS of each user in terms of the delay bound and overall packet loss probability. The global optimal solution is found in a symmetric scenario. A neural network was introduced to find an approximated optimal solution in general scenarios, where the QoS is ensured by using the property that the optimal solution should satisfy as the "supervision signal". Simulation results show that the learning-based solution performs the same as the optimal solution in the symmetric scenario, and can save around 40% bandwidth with respect to the state-of-the-art policy.