Efficient Neural Network Training via Forward and Backward Propagation Sparsification

Sparse training is a natural idea to accelerate the training speed of deep neural networks and save the memory usage, especially since large modern neural networks are significantly over-parameterized. However, most of the existing methods cannot achieve this goal in practice because the chain rule based gradient (w.r.t. structure parameters) estimators adopted by previous methods require dense computation at least in the backward propagation step. This paper solves this problem by proposing an efficient sparse training method with completely sparse forward and backward passes. We first formulate the training process as a continuous minimization problem under global sparsity constraint. We then separate the optimization process into two steps, corresponding to weight update and structure parameter update. For the former step, we use the conventional chain rule, which can be sparse via exploiting the sparse structure. For the latter step, instead of using the chain rule based gradient estimators as in existing methods, we propose a variance reduced policy gradient estimator, which only requires two forward passes without backward propagation, thus achieving completely sparse training. We prove that the variance of our gradient estimator is bounded. Extensive experimental results on real-world datasets demonstrate that compared to previous methods, our algorithm is much more effective in accelerating the training process, up to an order of magnitude faster.

Why Stable Learning Works? A Theory of Covariate Shift Generalization

Covariate shift generalization, a typical case in out-of-distribution (OOD) generalization, requires a good performance on the unknown testing distribution, which varies from the accessible training distribution in the form of covariate shift. Recently, stable learning algorithms have shown empirical effectiveness to deal with covariate shift generalization on several learning models involving regression algorithms and deep neural networks. However, the theoretical explanations for such effectiveness are still missing. In this paper, we take a step further towards the theoretical analysis of stable learning algorithms by explaining them as feature selection processes. We first specify a set of variables, named minimal stable variable set, that is minimal and optimal to deal with covariate shift generalization for common loss functions, including the mean squared loss and binary cross entropy loss. Then we prove that under ideal conditions, stable learning algorithms could identify the variables in this set. Further analysis on asymptotic properties and error propagation are also provided. These theories shed light on why stable learning works for covariate shift generalization.

Eigencurve: Optimal Learning Rate Schedule for SGD on Quadratic Objectives with Skewed Hessian Spectrums

Learning rate schedulers have been widely adopted in training deep neural networks. Despite their practical importance, there is a discrepancy between its practice and its theoretical analysis. For instance, it is not known what schedules of SGD achieve best convergence, even for simple problems such as optimizing quadratic objectives. So far, step decay has been one of the strongest candidates under this setup, which is proved to be nearly optimal with a $\mathcal{O}(\log T)$ gap. However, according to our analysis, this gap turns out to be $\Omega(\log T)$ in a wide range of settings, which throws the schedule optimality problem into an open question again. Towards answering this reopened question, in this paper, we propose Eigencurve, the first family of learning rate schedules that can achieve minimax optimal convergence rates (up to a constant) for SGD on quadratic objectives when the eigenvalue distribution of the underlying Hessian matrix is skewed. The condition is quite common in practice. Experimental results show that Eigencurve can significantly outperform step decay in image classification tasks on CIFAR-10, especially when the number of epochs is small. Moreover, the theory inspires two simple learning rate schedulers for practical applications that can approximate Eigencurve. For some problems, the optimal shape of the proposed schedulers resembles that of cosine decay, which sheds light to the success of cosine decay for such situations. For other situations, the proposed schedulers are superior to cosine decay.

Discovery-and-Selection: Towards Optimal Multiple Instance Learning for Weakly Supervised Object Detection

Weakly supervised object detection (WSOD) is a challenging task that requires simultaneously learn object classifiers and estimate object locations under the supervision of image category labels. A major line of WSOD methods roots in multiple instance learning which regards images as bags of instance and selects positive instances from each bag to learn the detector. However, a grand challenge emerges when the detector inclines to converge to discriminative parts of objects rather than the whole objects. In this paper, under the hypothesis that optimal solutions are included in local minima, we propose a discoveryand-selection approach fused with multiple instance learning (DS-MIL), which finds rich local minima and select optimal solutions from multiple local minima. To implement DS-MIL, an attention module is designed so that more context information can be captured by feature maps and more valuable proposals can be collected during training. With proposal candidates, a re-rank module is designed to select informative instances for object detector training. Experimental results on commonly used benchmarks show that our proposed DS-MIL approach can consistently improve the baselines, reporting state-of-the-art performance.

When is the Convergence Time of Langevin Algorithms Dimension Independent? A Composite Optimization Viewpoint

There has been a surge of works bridging MCMC sampling and optimization, with a specific focus on translating non-asymptotic convergence guarantees for optimization problems into the analysis of Langevin algorithms in MCMC sampling. A conspicuous distinction between the convergence analysis of Langevin sampling and that of optimization is that all known convergence rates for Langevin algorithms depend on the dimensionality of the problem, whereas the convergence rates for optimization are dimension-free for convex problems. Whether a dimension independent convergence rate can be achieved by Langevin algorithm is thus a long-standing open problem. This paper provides an affirmative answer to this problem for large classes of either Lipschitz or smooth convex problems with normal priors. By viewing Langevin algorithm as composite optimization, we develop a new analysis technique that leads to dimension independent convergence rates for such problems.

Feel-Good Thompson Sampling for Contextual Bandits and Reinforcement Learning

Thompson Sampling has been widely used for contextual bandit problems due to the flexibility of its modeling power. However, a general theory for this class of methods in the frequentist setting is still lacking. In this paper, we present a theoretical analysis of Thompson Sampling, with a focus on frequentist regret bounds. In this setting, we show that the standard Thompson Sampling is not aggressive enough in exploring new actions, leading to suboptimality in some pessimistic situations. A simple modification called Feel-Good Thompson Sampling, which favors high reward models more aggressively than the standard Thompson Sampling, is proposed to remedy this problem. We show that the theoretical framework can be used to derive Bayesian regret bounds for standard Thompson Sampling, and frequentist regret bounds for Feel-Good Thompson Sampling. It is shown that in both cases, we can reduce the bandit regret problem to online least squares regression estimation. For the frequentist analysis, the online least squares regression bound can be directly obtained using online aggregation techniques which have been well studied. The resulting bandit regret bound matches the minimax lower bound in the finite action case. Moreover, the analysis can be generalized to handle a class of linearly embeddable contextual bandit problems (which generalizes the popular linear contextual bandit model). The obtained result again matches the minimax lower bound. Finally we illustrate that the analysis can be extended to handle some MDP problems.

EllipseNet: Anchor-Free Ellipse Detection for Automatic Cardiac Biometrics in Fetal Echocardiography

As an important scan plane, four chamber view is routinely performed in both second trimester perinatal screening and fetal echocardiographic examinations. The biometrics in this plane including cardio-thoracic ratio (CTR) and cardiac axis are usually measured by sonographers for diagnosing congenital heart disease. However, due to the commonly existing artifacts like acoustic shadowing, the traditional manual measurements not only suffer from the low efficiency, but also with the inconsistent results depending on the operators' skills. In this paper, we present an anchor-free ellipse detection network, namely EllipseNet, which detects the cardiac and thoracic regions in ellipse and automatically calculates the CTR and cardiac axis for fetal cardiac biometrics in 4-chamber view. In particular, we formulate the network that detects the center of each object as points and regresses the ellipses' parameters simultaneously. We define an intersection-over-union loss to further regulate the regression procedure. We evaluate EllipseNet on clinical echocardiogram dataset with more than 2000 subjects. Experimental results show that the proposed framework outperforms several state-of-the-art methods. Source code will be available at https://git.openi.org.cn/capepoint/EllipseNet .

Feature Correlation Aggregation: on the Path to Better Graph Neural Networks

Prior to the introduction of Graph Neural Networks (GNNs), modeling and analyzing irregular data, particularly graphs, was thought to be the Achilles' heel of deep learning. The core concept of GNNs is to find a representation by recursively aggregating the representations of a central node and those of its neighbors. The core concept of GNNs is to find a representation by recursively aggregating the representations of a central node and those of its neighbor, and its success has been demonstrated by many GNNs' designs. However, most of them only focus on using the first-order information between a node and its neighbors. In this paper, we introduce a central node permutation variant function through a frustratingly simple and innocent-looking modification to the core operation of a GNN, namely the Feature cOrrelation aGgregation (FOG) module which learns the second-order information from feature correlation between a node and its neighbors in the pipeline. By adding FOG into existing variants of GNNs, we empirically verify this second-order information complements the features generated by original GNNs across a broad set of benchmarks. A tangible boost in performance of the model is observed where the model surpasses previous state-of-the-art results by a significant margin while employing fewer parameters. (e.g., 33.116% improvement on a real-world molecular dataset using graph convolutional networks).

G-DetKD: Towards General Distillation Framework for Object Detectors via Contrastive and Semantic-guided Feature Imitation

In this paper, we investigate the knowledge distillation (KD) strategy for object detection and propose an effective framework applicable to both homogeneous and heterogeneous student-teacher pairs. The conventional feature imitation paradigm introduces imitation masks to focus on informative foreground areas while excluding the background noises. However, we find that those methods fail to fully utilize the semantic information in all feature pyramid levels, which leads to inefficiency for knowledge distillation between FPN-based detectors. To this end, we propose a novel semantic-guided feature imitation technique, which automatically performs soft matching between feature pairs across all pyramid levels to provide the optimal guidance to the student. To push the envelop even further, we introduce contrastive distillation to effectively capture the information encoded in the relationship between different feature regions. Finally, we propose a generalized detection KD pipeline, which is capable of distilling both homogeneous and heterogeneous detector pairs. Our method consistently outperforms the existing detection KD techniques, and works when (1) components in the framework are used separately and in conjunction; (2) for both homogeneous and heterogenous student-teacher pairs and (3) on multiple detection benchmarks. With a powerful X101-FasterRCNN-Instaboost detector as the teacher, R50-FasterRCNN reaches 44.0% AP, R50-RetinaNet reaches 43.3% AP and R50-FCOS reaches 43.1% AP on COCO dataset.

Accelerating Edge Intelligence via Integrated Sensing and Communication

Realizing edge intelligence consists of sensing, communication, training, and inference stages. Conventionally, the sensing and communication stages are executed sequentially, which results in excessive amount of dataset generation and uploading time. This paper proposes to accelerate edge intelligence via integrated sensing and communication (ISAC). As such, the sensing and communication stages are merged so as to make the best use of the wireless signals for the dual purpose of dataset generation and uploading. However, ISAC also introduces additional interference between sensing and communication functionalities. To address this challenge, this paper proposes a classification error minimization formulation to design the ISAC beamforming and time allocation. Globally optimal solution is derived via the rank-1 guaranteed semidefinite relaxation, and performance analysis is performed to quantify the ISAC gain. Simulation results are provided to verify the effectiveness of the proposed ISAC scheme. Interestingly, it is found that when the sensing time dominates the communication time, ISAC is always beneficial. However, when the communication time dominates, the edge intelligence with ISAC scheme may not be better than that with the conventional scheme, since ISAC introduces harmful interference between the sensing and communication signals.