Label hierarchies are often available apriori as part of biological taxonomy or language datasets WordNet. Several works exploit these to learn hierarchy aware features in order to improve the classifier to make semantically meaningful mistakes while maintaining or reducing the overall error. In this paper, we propose a novel approach for learning Hierarchy Aware Features (HAF) that leverages classifiers at each level of the hierarchy that are constrained to generate predictions consistent with the label hierarchy. The classifiers are trained by minimizing a Jensen-Shannon Divergence with target soft labels obtained from the fine-grained classifiers. Additionally, we employ a simple geometric loss that constrains the feature space geometry to capture the semantic structure of the label space. HAF is a training time approach that improves the mistakes while maintaining top-1 error, thereby, addressing the problem of cross-entropy loss that treats all mistakes as equal. We evaluate HAF on three hierarchical datasets and achieve state-of-the-art results on the iNaturalist-19 and CIFAR-100 datasets. The source code is available at https://github.com/07Agarg/HAF
Neural schedulers based on deep reinforcement learning (DRL) have shown considerable potential for solving real-world resource allocation problems, as they have demonstrated significant performance gain in the domain of cluster computing. In this paper, we investigate the feasibility of neural schedulers for the domain of System-on-Chip (SoC) resource allocation through extensive experiments and comparison with non-neural, heuristic schedulers. The key finding is three-fold. First, neural schedulers designed for cluster computing domain do not work well for SoC due to i) heterogeneity of SoC computing resources and ii) variable action set caused by randomness in incoming jobs. Second, our novel neural scheduler technique, Eclectic Interaction Matching (EIM), overcomes the above challenges, thus significantly improving the existing neural schedulers. Specifically, we rationalize the underlying reasons behind the performance gain by the EIM-based neural scheduler. Third, we discover that the ratio of the average processing elements (PE) switching delay and the average PE computation time significantly impacts the performance of neural SoC schedulers even with EIM. Consequently, future neural SoC scheduler design must consider this metric as well as its implementation overhead for practical utility.
We study differentially private (DP) algorithms for smooth stochastic minimax optimization, with stochastic minimization as a byproduct. The holy grail of these settings is to guarantee the optimal trade-off between the privacy and the excess population loss, using an algorithm with a linear time-complexity in the number of training samples. We provide a general framework for solving differentially private stochastic minimax optimization (DP-SMO) problems, which enables the practitioners to bring their own base optimization algorithm and use it as a black-box to obtain the near-optimal privacy-loss trade-off. Our framework is inspired from the recently proposed Phased-ERM method [20] for nonsmooth differentially private stochastic convex optimization (DP-SCO), which exploits the stability of the empirical risk minimization (ERM) for the privacy guarantee. The flexibility of our approach enables us to sidestep the requirement that the base algorithm needs to have bounded sensitivity, and allows the use of sophisticated variance-reduced accelerated methods to achieve near-linear time-complexity. To the best of our knowledge, these are the first linear-time optimal algorithms, up to logarithmic factors, for smooth DP-SMO when the objective is (strongly-)convex-(strongly-)concave. Additionally, based on our flexible framework, we derive a new family of near-linear time algorithms for smooth DP-SCO with optimal privacy-loss trade-offs for a wider range of smoothness parameters compared to previous algorithms.
Reducing speckle and limiting the variations of the physical parameters in Synthetic Aperture Radar (SAR) images is often a key-step to fully exploit the potential of such data. Nowadays, deep learning approaches produce state of the art results in single-image SAR restoration. Nevertheless, huge multi-temporal stacks are now often available and could be efficiently exploited to further improve image quality. This paper explores two fast strategies employing a single-image despeckling algorithm, namely SAR2SAR, in a multi-temporal framework. The first one is based on Quegan filter and replaces the local reflectivity pre-estimation by SAR2SAR. The second one uses SAR2SAR to suppress speckle from a ratio image encoding the multi-temporal information under the form of a "super-image", i.e. the temporal arithmetic mean of a time series. Experimental results on Sentinel-1 GRD data show that these two multi-temporal strategies provide improved filtering results while adding a limited computational cost.
This contribution introduces a novel signal extrapolation algorithm and its application to image error concealment. The signal extrapolation is carried out by iteratively generating a model of the signal suffering from distortion. Thereby, the model results from a weighted superposition of two-dimensional basis functions whereas in every iteration step a set of these is selected and the approximation residual is projected onto the subspace they span. The algorithm is an improvement to the Frequency Selective Extrapolation that has proven to be an effective method for concealing lost or distorted image regions. Compared to this algorithm, the novel algorithm is able to reduce the processing time by a factor larger than three, by still preserving the very high extrapolation quality.
Recommendation systems play an important role in today's digital world. They have found applications in various applications such as music platforms, e.g., Spotify, and movie streaming services, e.g., Netflix. Less research effort has been devoted to physical exercise recommendation systems. Sedentary lifestyles have become the major driver of several diseases as well as healthcare costs. In this paper, we develop a recommendation system for daily exercise activities to users based on their history, profile and similar users. The developed recommendation system uses a deep recurrent neural network with user-profile attention and temporal attention mechanisms. Moreover, exercise recommendation systems are significantly different from streaming recommendation systems in that we are not able to collect click feedback from the participants in exercise recommendation systems. Thus, we propose a real-time, expert-in-the-loop active learning procedure. The active learners calculate the uncertainty of the recommender at each time step for each user and ask an expert for a recommendation when the certainty is low. In this paper, we derive the probability distribution function of marginal distance, and use it to determine when to ask experts for feedback. Our experimental results on a mHealth dataset show improved accuracy after incorporating the real-time active learner with the recommendation system.
Deep Neural Network (DNN) models are usually trained sequentially from one layer to another, which causes forward, backward and update locking's problems, leading to poor performance in terms of training time. The existing parallel strategies to mitigate these problems provide suboptimal runtime performance. In this work, we have proposed a novel layer-wise partitioning and merging, forward and backward pass parallel framework to provide better training performance. The novelty of the proposed work consists of 1) a layer-wise partition and merging model which can minimise communication overhead between devices without the memory cost of existing strategies during the training process; 2) a forward pass and backward pass parallelisation and optimisation to address the update locking problem and minimise the total training cost. The experimental evaluation on real use cases shows that the proposed method outperforms the state-of-the-art approaches in terms of training speed; and achieves almost linear speedup without compromising the accuracy performance of the non-parallel approach.
Sports game summarization aims to generate sports news based on real-time commentaries. The task has attracted wide research attention but is still under-explored probably due to the lack of corresponding English datasets. Therefore, in this paper, we release GOAL, the first English sports game summarization dataset. Specifically, there are 103 commentary-news pairs in GOAL, where the average lengths of commentaries and news are 2724.9 and 476.3 words, respectively. Moreover, to support the research in the semi-supervised setting, GOAL additionally provides 2,160 unlabeled commentary documents. Based on our GOAL, we build and evaluate several baselines, including extractive and abstractive baselines. The experimental results show the challenges of this task still remain. We hope our work could promote the research of sports game summarization. The dataset has been released at https://github.com/krystalan/goal.
Photoacoustic (PA) imaging is an emerging hybrid imaging modality that combines rich optical spectroscopic contrast and high ultrasonic resolution and thus holds tremendous promise for a wide range of pre-clinical and clinical applications. Compact and affordable light sources such as light-emitting diodes (LEDs) and laser diodes (LDs) are promising alternatives to bulky and expensive solid-state laser systems that are commonly used as PA light sources. These could accelerate the clinical translation of PA technology. However, PA signals generated with these light sources are readily degraded by noise due to the low optical fluence, leading to decreased signal-to-noise ratio (SNR) in PA images. In this work, a spatiotemporal singular value decomposition (SVD) based PA denoising method was investigated for these light sources that usually have low fluence and high repetition rates. The proposed method leverages both spatial and temporal correlations between radiofrequency (RF) data frames. Validation was performed on simulations and in vivo PA data acquired from human fingers (2D) and forearm (3D) using a LED-based system. Spatiotemporal SVD greatly enhanced the PA signals of blood vessels corrupted by noise while preserving a high temporal resolution to slow motions, improving the SNR of in vivo PA images by 1.1, 0.7, and 1.9 times compared to single frame-based wavelet denoising, averaging across 200 frames, and single frame without denoising, respectively. The proposed method demonstrated a processing time of around 50 \mus per frame with SVD acceleration and GPU. Thus, spatiotemporal SVD is well suited to PA imaging systems with low-energy excitation light sources for real-time in vivo applications.
Changes, planned or unexpected, are common during the execution of real-life processes. Detecting these changes is a must for optimizing the performance of organizations running such processes. Most of the algorithms present in the state-of-the-art focus on the detection of sudden changes, leaving aside other types of changes. In this paper, we will focus on the automatic detection of gradual drifts, a special type of change, in which the cases of two models overlap during a period of time. The proposed algorithm relies on conformance checking metrics to carry out the automatic detection of the changes, performing also a fully automatic classification of these changes into sudden or gradual. The approach has been validated with a synthetic dataset consisting of 120 logs with different distributions of changes, getting better results in terms of detection and classification accuracy, delay and change region overlapping than the main state-of-the-art algorithms.