This paper introduces an open source platform for rapid development of computer vision applications. The platform puts the efficient data development at the center of the machine learning development process, integrates active learning methods, data and model version control, and uses concepts such as projects to enable fast iteration of multiple task specific datasets in parallel. We make it an open platform by abstracting the development process into core states and operations, and design open APIs to integrate third party tools as implementations of the operations. This open design reduces the development cost and adoption cost for ML teams with existing tools. At the same time, the platform supports recording project development history, through which successful projects can be shared to further boost model production efficiency on similar tasks. The platform is open source and is already used internally to meet the increasing demand from custom real world computer vision applications.
Exploration and analysis of massive datasets has recently generated increasing interest in the research and development communities. It has long been a recognized problem that many datasets contain significant levels of missing numerical data. We introduce a mathematically principled stochastic optimization imputation method based on the theory of Kriging. This is shown to be a powerful method for imputation. However, its computational effort and potential numerical instabilities produce costly and/or unreliable predictions, potentially limiting its use on large scale datasets. In this paper, we apply a recently developed multi-level stochastic optimization approach to the problem of imputation in massive medical records. The approach is based on computational applied mathematics techniques and is highly accurate. In particular, for the Best Linear Unbiased Predictor (BLUP) this multi-level formulation is exact, and is also significantly faster and more numerically stable. This permits practical application of Kriging methods to data imputation problems for massive datasets. We test this approach on data from the National Inpatient Sample (NIS) data records, Healthcare Cost and Utilization Project (HCUP), Agency for Healthcare Research and Quality. Numerical results show the multi-level method significantly outperforms current approaches and is numerically robust. In particular, it has superior accuracy as compared with methods recommended in the recent report from HCUP on the important problem of missing data, which could lead to sub-optimal and poorly based funding policy decisions. In comparative benchmark tests it is shown that the multilevel stochastic method is significantly superior to recommended methods in the report, including Predictive Mean Matching (PMM) and Predicted Posterior Distribution (PPD), with up to 75% reductions in error.
There has been a recent surge of research interest in attacking the problem of social relation inference based on images. Existing works classify social relations mainly by creating complicated graphs of human interactions, or learning the foreground and/or background information of persons and objects, but ignore holistic scene context. The holistic scene refers to the functionality of a place in images, such as dinning room, playground and office. In this paper, by mimicking human understanding on images, we propose an approach of \textbf{PR}actical \textbf{I}nference in \textbf{S}ocial r\textbf{E}lation (PRISE), which concisely learns interactive features of persons and discriminative features of holistic scenes. Technically, we develop a simple and fast relational graph convolutional network to capture interactive features of all persons in one image. To learn the holistic scene feature, we elaborately design a contrastive learning task based on image scene classification. To further boost the performance in social relation inference, we collect and distribute a new large-scale dataset, which consists of about 240 thousand unlabeled images. The extensive experimental results show that our novel learning framework significantly beats the state-of-the-art methods, e.g., PRISE achieves 6.8$\%$ improvement for domain classification in PIPA dataset.
Approximate Nearest neighbor search (ANNS) plays a crucial role in information retrieval, which has a wide range of application scenarios. Therefore, during past several years, a lot of fast ANNS approaches have been proposed. Among these approaches, graph-based methods are one of the most popular type, as they have shown attractive theoretical guarantees and low query latency. In this paper, we propose a learnable compression network with transformer (LCNT), which projects feature vectors from high dimensional space onto low dimensional space, while preserving neighbor relationship. The proposed model can be generalized to existing graph-based methods to accelerate the process of building indexing graph and further reduce query latency. Specifically, the proposed LCNT contains two major parts, projection part and harmonizing part. In the projection part, input vectors are projected into a sequence of subspaces via multi channel sparse projection network. In the harmonizing part, a modified Transformer network is employed to harmonize features in subspaces and combine them to get a new feature. To evaluate the effectiveness of the proposed model, we conduct experiments on two million-scale databases, GIST1M and Deep1M. Experimental results show that the proposed model can improve the speed of building indexing graph to 2-3 times its original speed without sacrificing accuracy significantly. The query latency is reduced by a factor of 1.3 to 2.0. In addition, the proposed model can also be combined with other popular quantization methods.
Deep neural networks (DNNs) are vulnerable to adversarial noise. A range of adversarial defense techniques have been proposed to mitigate the interference of adversarial noise, among which the input pre-processing methods are scalable and show great potential to safeguard DNNs. However, pre-processing methods may suffer from the robustness degradation effect, in which the defense reduces rather than improving the adversarial robustness of a target model in a white-box setting. A potential cause of this negative effect is that adversarial training examples are static and independent to the pre-processing model. To solve this problem, we investigate the influence of full adversarial examples which are crafted against the full model, and find they indeed have a positive impact on the robustness of defenses. Furthermore, we find that simply changing the adversarial training examples in pre-processing methods does not completely alleviate the robustness degradation effect. This is due to the adversarial risk of the pre-processed model being neglected, which is another cause of the robustness degradation effect. Motivated by above analyses, we propose a method called Joint Adversarial Training based Pre-processing (JATP) defense. Specifically, we formulate a feature similarity based adversarial risk for the pre-processing model by using full adversarial examples found in a feature space. Unlike standard adversarial training, we only update the pre-processing model, which prompts us to introduce a pixel-wise loss to improve its cross-model transferability. We then conduct a joint adversarial training on the pre-processing model to minimize this overall risk. Empirical results show that our method could effectively mitigate the robustness degradation effect across different target models in comparison to previous state-of-the-art approaches.
Many popular learning-rate schedules for deep neural networks combine a decaying trend with local perturbations that attempt to escape saddle points and bad local minima. We derive convergence guarantees for bandwidth-based step-sizes, a general class of learning-rates that are allowed to vary in a banded region. This framework includes cyclic and non-monotonic step-sizes for which no theoretical guarantees were previously known. We provide worst-case guarantees for SGD on smooth non-convex problems under several bandwidth-based step sizes, including stagewise $1/\sqrt{t}$ and the popular step-decay (constant and then drop by a constant), which is also shown to be optimal. Moreover, we show that its momentum variant (SGDM) converges as fast as SGD with the bandwidth-based step-decay step-size. Finally, we propose some novel step-size schemes in the bandwidth-based family and verify their efficiency on several deep neural network training tasks.
Deep neural networks (DNNs) are vulnerable to adversarial noise. Preprocessing based defenses could largely remove adversarial noise by processing inputs. However, they are typically affected by the error amplification effect, especially in the front of continuously evolving attacks. To solve this problem, in this paper, we propose to remove adversarial noise by implementing a self-supervised adversarial training mechanism in a class activation feature space. To be specific, we first maximize the disruptions to class activation features of natural examples to craft adversarial examples. Then, we train a denoising model to minimize the distances between the adversarial examples and the natural examples in the class activation feature space. Empirical evaluations demonstrate that our method could significantly enhance adversarial robustness in comparison to previous state-of-the-art approaches, especially against unseen adversarial attacks and adaptive attacks.
The convergence of stochastic gradient descent is highly dependent on the step-size, especially on non-convex problems such as neural network training. Step decay step-size schedules (constant and then cut) are widely used in practice because of their excellent convergence and generalization qualities, but their theoretical properties are not yet well understood. We provide the convergence results for step decay in the non-convex regime, ensuring that the gradient norm vanishes at an $\mathcal{O}(\ln T/\sqrt{T})$ rate. We also provide the convergence guarantees for general (possibly non-smooth) convex problems, ensuring an $\mathcal{O}(\ln T/\sqrt{T})$ convergence rate. Finally, in the strongly convex case, we establish an $\mathcal{O}(\ln T/T)$ rate for smooth problems, which we also prove to be tight, and an $\mathcal{O}(\ln^2 T /T)$ rate without the smoothness assumption. We illustrate the practical efficiency of the step decay step-size in several large scale deep neural network training tasks.
Gaussian noise injections (GNIs) are a family of simple and widely-used regularisation methods for training neural networks, where one injects additive or multiplicative Gaussian noise to the network activations at every iteration of the optimisation algorithm, which is typically chosen as stochastic gradient descent (SGD). In this paper we focus on the so-called `implicit effect' of GNIs, which is the effect of the injected noise on the dynamics of SGD. We show that this effect induces an asymmetric heavy-tailed noise on SGD gradient updates. In order to model this modified dynamics, we first develop a Langevin-like stochastic differential equation that is driven by a general family of asymmetric heavy-tailed noise. Using this model we then formally prove that GNIs induce an `implicit bias', which varies depending on the heaviness of the tails and the level of asymmetry. Our empirical results confirm that different types of neural networks trained with GNIs are well-modelled by the proposed dynamics and that the implicit effect of these injections induces a bias that degrades the performance of networks.
We present a generalisation of Rosenblatt's traditional perceptron learning algorithm to the class of proximal activation functions and demonstrate how this generalisation can be interpreted as an incremental gradient method applied to a novel energy function. This novel energy function is based on a generalised Bregman distance, for which the gradient with respect to the weights and biases does not require the differentiation of the activation function. The interpretation as an energy minimisation algorithm paves the way for many new algorithms, of which we explore a novel variant of the iterative soft-thresholding algorithm for the learning of sparse perceptrons.