Driver drowsiness increases crash risk, leading to substantial road trauma each year. Drowsiness detection methods have received considerable attention, but few studies have investigated the implementation of a detection approach on a mobile phone. Phone applications reduce the need for specialised hardware and hence, enable a cost-effective roll-out of the technology across the driving population. While it has been shown that three-dimensional (3D) operations are more suitable for spatiotemporal feature learning, current methods for drowsiness detection commonly use frame-based, multi-step approaches. However, computationally expensive techniques that achieve superior results on action recognition benchmarks (e.g. 3D convolutions, optical flow extraction) create bottlenecks for real-time, safety-critical applications on mobile devices. Here, we show how depthwise separable 3D convolutions, combined with an early fusion of spatial and temporal information, can achieve a balance between high prediction accuracy and real-time inference requirements. In particular, increased accuracy is achieved when assessment requires motion information, for example, when sunglasses conceal the eyes. Further, a custom TensorFlow-based smartphone application shows the true impact of various approaches on inference times and demonstrates the effectiveness of real-time monitoring based on out-of-sample data to alert a drowsy driver. Our model is pre-trained on ImageNet and Kinetics and fine-tuned on a publicly available Driver Drowsiness Detection dataset. Fine-tuning on large naturalistic driving datasets could further improve accuracy to obtain robust in-vehicle performance. Overall, our research is a step towards practical deep learning applications, potentially preventing micro-sleeps and reducing road trauma.

--Existing algorithms aiming to learn a binary classifier from positive (P) and unlabeled (U) data require estimating the class prior or label noise ahead of building a classification model. However, the estimation and classifier learning are normally conducted in a pipeline instead of being jointly optimized. In this paper, we propose to alternatively train the two steps using reinforcement learning. Our proposal adopts a policy network to adaptively make assumptions on the labels of unlabeled data, while a classifier is built upon the output of the policy network and provides rewards to learn a better policy. The dynamic and interactive training between the policy maker and the classifier can exploit the unlabeled data in a more effective manner and yield a significant improvement in terms of classification performance. Furthermore, we present two different approaches to represent the actions taken by the policy. The first approach considers continuous actions as soft labels, while the other uses discrete actions as hard assignment of labels for unlabeled examples. We validate the effectiveness of the proposed method on two public benchmark datasets as well as one e-commerce dataset. The results show that the proposed method is able to consistently outperform state-of-the-art methods in various settings. PU learning refers to the problem of learning from a dataset where only a subset of examples are positively labeled and the rest are not annotated at all. It is a critical task due to its prevalence in various real-world applications [1], [2], [3]. In many common situations only positive data are available, for instance, an e-commerce website may only record users who have clicked on advertisements or purchased items. Meanwhile, it is not possible to simply assume that unlabeled instances are negative.

The main goal of this work is equipping convex and nonconvex problems with Barzilai-Borwein (BB) step size. With the adaptivity of BB step sizes granted, they can fail when the objective function is not strongly convex. To overcome this challenge, the key idea here is to bridge (non)convex problems and strongly convex ones via regularization. The proposed regularization schemes are \textit{simple} yet effective. Wedding the BB step size with a variance reduction method, known as SARAH, offers a free lunch compared with vanilla SARAH in convex problems. The convergence of BB step sizes in nonconvex problems is also established and its complexity is no worse than other adaptive step sizes such as AdaGrad. As a byproduct, our regularized SARAH methods for convex functions ensure that the complexity to find $\mathbb{E}[\| \nabla f(\mathbf{x}) \|^2]\leq \epsilon$ is ${\cal O}\big( (n+\frac{1}{\sqrt{\epsilon}})\ln{\frac{1}{\epsilon}}\big)$, improving $\epsilon$ dependence over existing results. Numerical tests further validate the merits of proposed approaches.

The adaptive momentum method (AdaMM), which uses past gradients to update descent directions and learning rates simultaneously, has become one of the most popular first-order optimization methods for solving machine learning problems. However, AdaMM is not suited for solving black-box optimization problems, where explicit gradient forms are difficult or infeasible to obtain. In this paper, we propose a zeroth-order AdaMM (ZO-AdaMM) algorithm, that generalizes AdaMM to the gradient-free regime. We show that the convergence rate of ZO-AdaMM for both convex and nonconvex optimization is roughly a factor of $O(\sqrt{d})$ worse than that of the first-order AdaMM algorithm, where $d$ is problem size. In particular, we provide a deep understanding on why Mahalanobis distance matters in convergence of ZO-AdaMM and other AdaMM-type methods. As a byproduct, our analysis makes the first step toward understanding adaptive learning rate methods for nonconvex constrained optimization. Furthermore, we demonstrate two applications, designing per-image and universal adversarial attacks from black-box neural networks, respectively. We perform extensive experiments on ImageNet and empirically show that ZO-AdaMM converges much faster to a solution of high accuracy compared with $6$ state-of-the-art ZO optimization methods.

Many machine learning tasks such as clustering, classification, and dataset search benefit from embedding data points in a space where distances reflect notions of relative similarity as perceived by humans. A common way to construct such an embedding is to request triplet similarity queries to an oracle, comparing two objects with respect to a reference. This work generalizes triplet queries to tuple queries of arbitrary size that ask an oracle to rank multiple objects against a reference, and introduces an efficient and robust adaptive selection method called InfoTuple that uses a novel approach to mutual information maximization. We show that the performance of InfoTuple at various tuple sizes exceeds that of the state-of-the-art adaptive triplet selection method on synthetic tests and new human response datasets, and empirically demonstrate the significant gains in efficiency and query consistency achieved by querying larger tuples instead of triplets.

Existing studies of the Multi Agent Multi Armed Bandit (MAMAB) problem, with the exception of a very few, consider the case where the agents observe their neighbors according to a static network graph. They also mostly rely on a running consensus for the estimation of the option rewards. Two of the exceptions consider a problem where agents observe instantaneous rewards and actions of their neighbors through an iid ER graph process based communication strategy. In this paper we propose a UCB based option allocation rule that guarantees logarithmic regret even if the graph depends on the history of choices made by the agents. The paper also proposes a novel communication strategy that significantly outperforms the iid ER graph based communication strategy. In both the ER graph and the dependent graph strategy, the regret is shown to depend on the connectivity of the graph in a particularly interesting way where there exists an optimal connectivity of the graph that is less than the full connectivity of the graph.

Addressing the non-uniform missing mechanism of rating feedback is critical to build a well-performing recommeder in the real-world systems. To tackle the challenging issue, we first define an ideal loss function that should be optimized to achieve the goal of recommendation. Then, we derive the generalization error bound of the ideal loss that alleviates the variance and the misspecification problems of the previous propensity-based methods. We further propose a meta-learning method minimizing the bound. Empirical evaluation using real-world datasets validates the theoretical findings and demonstrates the practical advantages of the proposed upper bound minimization approach.

Conventional hardware-friendly quantization methods, such as fixed-point or integer, tend to perform poorly at very low word sizes as their shrinking dynamic ranges cannot adequately capture the wide data distributions commonly seen in sequence transduction models. We present AdaptivFloat, a floating-point inspired number representation format for deep learning that dynamically maximizes and optimally clips its available dynamic range, at a layer granularity, in order to create faithful encoding of neural network parameters. AdaptivFloat consistently produces higher inference accuracies compared to block floating-point, uniform, IEEE-like float or posit encodings at very low precision ($\leq$ 8-bit) across a diverse set of state-of-the-art neural network topologies. And notably, AdaptivFloat is seen surpassing baseline FP32 performance by up to +0.3 in BLEU score and -0.75 in word error rate at weight bit widths that are $\leq$ 8-bit. Experimental results on a deep neural network (DNN) hardware accelerator, exploiting AdaptivFloat logic in its computational datapath, demonstrate per-operation energy and area that is 0.9$\times$ and 1.14$\times$, respectively, that of equivalent bit width integer-based accelerator variants.

Predicting delayed outcomes is an important problem in recommender systems (e.g., if customers will finish reading an ebook). We formalize the problem as an adversarial, delayed online learning problem and consider how a proxy for the delayed outcome (e.g., if customers read a third of the book in 24 hours) can help minimize regret, even though the proxy is not available when making a prediction. Motivated by our regret analysis, we propose two neural network architectures: Factored Forecaster (FF) which is ideal if the proxy is informative of the outcome in hindsight, and Residual Factored Forecaster (RFF) that is robust to a non-informative proxy. Experiments on two real-world datasets for predicting human behavior show that RFF outperforms both FF and a direct forecaster that does not make use of the proxy. Our results suggest that exploiting proxies by factorization is a promising way to mitigate the impact of long delays in human-behavior prediction tasks.

This paper further defines the class of fragment constructive heuristics used to compute feasible solutions for the Traveling Salesman Problem into arc-greedy and node-greedy subclasses. Since these subclasses of heuristics can create subtours, two known methodologies for subtour elimination on symmetric instances are reviewed and are expanded to cover asymmetric problem instances. This paper introduces a third novel methodology, the Greedy Tracker, and compares it to both known methodologies. Computational results are generated across multiple symmetric and asymmetric instances. The results demonstrate the Greedy Tracker is the fastest method for preventing subtours for instances below 400 nodes. A distinction between fragment constructive heuristics and the subtour elimination methodology used to ensure the feasibility of resulting solutions enables the introduction of a new node-greedy fragment heuristic called Ordered Greedy.