Learn in-situ is a growing trend for Edge AI. Training deep neural network (DNN) on edge devices is challenging because both energy and memory are constrained. Low precision training helps to reduce the energy cost of a single training iteration, but that does not necessarily translate to energy savings for the whole training process, because low precision could slows down the convergence rate. One evidence is that most works for low precision training keep an fp32 copy of the model during training, which in turn imposes memory requirements on edge devices. In this work we propose Adaptive Precision Training. It is able to save both total training energy cost and memory usage at the same time. We use model of the same precision for both forward and backward pass in order to reduce memory usage for training. Through evaluating the progress of training, APT allocates layer-wise precision dynamically so that the model learns quicker for longer time. APT provides an application specific hyper-parameter for users to play trade-off between training energy cost, memory usage and accuracy. Experiment shows that APT achieves more than 50% saving on training energy and memory usage with limited accuracy loss. 20% more savings of training energy and memory usage can be achieved in return for a 1% sacrifice in accuracy loss.
We consider time-domain digital backpropagation with chromatic dispersion filters jointly optimized and quantized using machine-learning techniques. Compared to the baseline implementations, we show improved BER performance and >40% power dissipation reductions in 28-nm CMOS.
New deep-learning architectures are created every year, achieving state-of-the-art results in image recognition and leading to the belief that, in a few years, complex tasks such as sign language translation will be considerably easier, serving as a communication tool for the hearing-impaired community. On the other hand, these algorithms still need a lot of data to be trained and the dataset creation process is expensive, time-consuming, and slow. Thereby, this work aims to investigate techniques of digital image processing and machine learning that can be used to create a sign language dataset effectively. We argue about data acquisition, such as the frames per second rate to capture or subsample the videos, the background type, preprocessing, and data augmentation, using convolutional neural networks and object detection to create an image classifier and comparing the results based on statistical tests. Different datasets were created to test the hypotheses, containing 14 words used daily and recorded by different smartphones in the RGB color system. We achieved an accuracy of 96.38% on the test set and 81.36% on the validation set containing more challenging conditions, showing that 30 FPS is the best frame rate subsample to train the classifier, geometric transformations work better than intensity transformations, and artificial background creation is not effective to model generalization. These trade-offs should be considered in future work as a cost-benefit guideline between computational cost and accuracy gain when creating a dataset and training a sign recognition model.
This paper bridges discrete and continuous optimization approaches for decomposable submodular function minimization, in both the standard and parametric settings. We provide improved running times for this problem by reducing it to a number of calls to a maximum flow oracle. When each function in the decomposition acts on $O(1)$ elements of the ground set $V$ and is polynomially bounded, our running time is up to polylogarithmic factors equal to that of solving maximum flow in a sparse graph with $O(\vert V \vert)$ vertices and polynomial integral capacities. We achieve this by providing a simple iterative method which can optimize to high precision any convex function defined on the submodular base polytope, provided we can efficiently minimize it on the base polytope corresponding to the cut function of a certain graph that we construct. We solve this minimization problem by lifting the solutions of a parametric cut problem, which we obtain via a new efficient combinatorial reduction to maximum flow. This reduction is of independent interest and implies some previously unknown bounds for the parametric minimum $s,t$-cut problem in multiple settings.
Recovering expressive 3D human pose and mesh from in-the-wild images is greatly challenging due to the absence of the training data. Several optimization-based methods have been used to obtain pseudo-groundtruth (GT) 3D poses and meshes from GT 2D poses. However, they often produce bad ones with long running time because their frameworks are optimized on each sample only using 2D supervisions in a sequential way. To overcome the limitations, we present NeuralAnnot, a neural annotator that learns to construct in-the-wild expressive 3D human pose and mesh training sets. Our NeuralAnnot is trained on a large number of samples by 2D supervisions from a target in-the-wild dataset and 3D supervisions from auxiliary datasets with GT 3D poses in a parallel way. We show that our NeuralAnnot produces far better 3D pseudo-GTs with much shorter running time than the optimization-based methods, and the newly obtained training set brings great performance gain. The newly obtained training sets and codes will be publicly available.
Multitarget Tracking (MTT) is the problem of tracking the states of an unknown number of objects using noisy measurements, with important applications to autonomous driving, surveillance, robotics, and others. In the model-based Bayesian setting, there are conjugate priors that enable us to express the multi-object posterior in closed form, which could theoretically provide Bayes-optimal estimates. However, the posterior involves a super-exponential growth of the number of hypotheses over time, forcing state-of-the-art methods to resort to approximations for remaining tractable, which can impact their performance in complex scenarios. Model-free methods based on deep-learning provide an attractive alternative, as they can in principle learn the optimal filter from data, but to the best of our knowledge were never compared to current state-of-the-art Bayesian filters, specially not in contexts where accurate models are available. In this paper, we propose a high-performing deep-learning method for MTT based on the Transformer architecture and compare it to two state-of-the-art Bayesian filters, in a setting where we assume the correct model is provided. Although this gives an edge to the model-based filters, it also allows us to generate unlimited training data. We show that the proposed model outperforms state-of-the-art Bayesian filters in complex scenarios, while macthing their performance in simpler cases, which validates the applicability of deep-learning also in the model-based regime. The code for all our implementations is made available at (github link to be provided).
Over the past years, extensive research has been dedicated to developing robust platforms and data-driven dialogue models to support long-term human-robot interactions. However, little is known about how people's perception of robots and engagement with them develop over time and how these can be accurately assessed through implicit and continuous measurement techniques. In this paper, we investigate this by involving participants in three interaction sessions with multiple days of zero exposure in between. Each session consists of a joint task with a robot as well as two short social chats with it before and after the task. We measure participants' gaze patterns with a wearable eye-tracker and gauge their perception of the robot and engagement with it and the joint task using questionnaires. Results disclose that aversion of gaze in a social chat is an indicator of a robot's uncanniness and that the more people gaze at the robot in a joint task, the worse they perform. In contrast with most HRI literature, our results show that gaze towards an object of shared attention, rather than gaze towards a robotic partner, is the most meaningful predictor of engagement in a joint task. Furthermore, the analyses of long-term gaze patterns disclose that people's mutual gaze in a social chat develops congruently with their perceptions of the robot over time. These are key findings for the HRI community as they entail that gaze behavior can be used as an implicit measure of people's perception of robots in a social chat and of their engagement and task performance in a joint task.
Effective and real-time eyeblink detection is of wide-range applications, such as deception detection, drive fatigue detection, face anti-spoofing, etc. Although numerous of efforts have already been paid, most of them focus on addressing the eyeblink detection problem under the constrained indoor conditions with the relative consistent subject and environment setup. Nevertheless, towards the practical applications eyeblink detection in the wild is more required, and of greater challenges. However, to our knowledge this has not been well studied before. In this paper, we shed the light to this research topic. A labelled eyeblink in the wild dataset (i.e., HUST-LEBW) of 673 eyeblink video samples (i.e., 381 positives, and 292 negatives) is first established by us. These samples are captured from the unconstrained movies, with the dramatic variation on human attribute, human pose, illumination condition, imaging configuration, etc. Then, we formulate eyeblink detection task as a spatial-temporal pattern recognition problem. After locating and tracking human eye using SeetaFace engine and KCF tracker respectively, a modified LSTM model able to capture the multi-scale temporal information is proposed to execute eyeblink verification. A feature extraction approach that reveals appearance and motion characteristics simultaneously is also proposed. The experiments on HUST-LEBW reveal the superiority and efficiency of our approach. It also verifies that, the existing eyeblink detection methods cannot achieve satisfactory performance in the wild.
In ophthalmology, early fundus screening is an economic and effective way to prevent blindness caused by ophthalmic diseases. Clinically, due to the lack of medical resources, manual diagnosis is time-consuming and may delay the condition. With the development of deep learning, some researches on ophthalmic diseases have achieved good results, however, most of them are just based on one disease. During fundus screening, ophthalmologists usually give diagnoses of multi-disease on binocular fundus image, so we release a dataset with 8 diseases to meet the real medical scene, which contains 10,000 fundus images from both eyes of 5,000 patients. We did some benchmark experiments on it through some state-of-the-art deep neural networks. We found simply increasing the scale of network cannot bring good results for multi-disease classification, and a well-structured feature fusion method combines characteristics of multi-disease is needed. Through this work, we hope to advance the research of related fields.
We study the problem of combinatorial multi-armed bandits (CMAB) under strategic manipulations of rewards, where each arm can modify the emitted reward signals for its own interest. Our setting elaborates a more realistic model of adaptive arms that imposes relaxed assumptions compared to adversarial corruptions and adversarial attacks. Algorithms designed under strategic arms gain robustness in real applications while avoiding being overcautious and hampering the performance. We bridge the gap between strategic manipulations and adversarial attacks by investigating the optimal colluding strategy among arms under the MAB problem. We then propose a strategic variant of the combinatorial UCB algorithm, which has a regret of at most $O(m\log T + m B_{max})$ under strategic manipulations, where $T$ is the time horizon, $m$ is the number of arms, and $B_{max}$ is the maximum budget. We further provide lower bounds on the strategic budgets for attackers to incur certain regret of the bandit algorithm. Extensive experiments corroborate our theoretical findings on robustness and regret bounds, in a variety of regimes of manipulation budgets.