Regularization of Deep Neural Networks (DNNs) for the sake of improving their generalization capability is important and challenging. The development in this line benefits theoretical foundation of DNNs and promotes their usability in different areas of artificial intelligence. In this paper, we investigate the role of Rademacher complexity in improving generalization of DNNs and propose a novel regularizer rooted in Local Rademacher Complexity (LRC). While Rademacher complexity is well known as a distribution-free complexity measure of function class that help boost generalization of statistical learning methods, extensive study shows that LRC, its counterpart focusing on a restricted function class, leads to sharper convergence rates and potential better generalization given finite training sample. Our LRC based regularizer is developed by estimating the complexity of the function class centered at the minimizer of the empirical loss of DNNs. Experiments on various types of network architecture demonstrate the effectiveness of LRC regularization in improving generalization. Moreover, our method features the state-of-the-art result on the CIFAR-$10$ dataset with network architecture found by neural architecture search.
Transfer learning through fine-tuning a pre-trained neural network with an extremely large dataset, such as ImageNet, can significantly accelerate training while the accuracy is frequently bottlenecked by the limited dataset size of the new target task. To solve the problem, some regularization methods, constraining the outer layer weights of the target network using the starting point as references (SPAR), have been studied. In this paper, we propose a novel regularized transfer learning framework DELTA, namely DEep Learning Transfer using Feature Map with Attention. Instead of constraining the weights of neural network, DELTA aims to preserve the outer layer outputs of the target network. Specifically, in addition to minimizing the empirical loss, DELTA intends to align the outer layer outputs of two networks, through constraining a subset of feature maps that are precisely selected by attention that has been learned in an supervised learning manner. We evaluate DELTA with the state-of-the-art algorithms, including L2 and L2-SP. The experiment results show that our proposed method outperforms these baselines with higher accuracy for new tasks.
We interpret the variational inference of the Stochastic Gradient Descent (SGD) as minimizing a new potential function named the \textit{quasi-potential}. We analytically construct the quasi-potential function in the case when the loss function is convex and admits only one global minimum point. We show in this case that the quasi-potential function is related to the noise covariance structure of SGD via a partial differential equation of Hamilton-Jacobi type. This relation helps us to show that anisotropic noise leads to faster escape than isotropic noise. We then consider the dynamics of SGD in the case when the loss function is non-convex and admits several different local minima. In this case, we demonstrate an example that shows how the noise covariance structure plays a role in "implicit regularization", a phenomenon in which SGD favors some particular local minimum points. This is done through the relation between the noise covariance structure and the quasi-potential function. Our analysis is based on Large Deviations Theory (LDT), and they are validated by numerical experiments.
Deep transfer learning has acquired significant research interest. It makes use of pre-trained models that are learned from a source domain, and utilizes these models for the tasks in a target domain. Model-based deep transfer learning is arguably the most frequently used method. However, very little work has been devoted to enhancing deep transfer learning by focusing on the influence of data. In this work, we propose an instance-based approach to improve deep transfer learning in target domain. Specifically, we choose a pre-trained model which is learned from a source domain, and utilize this model to estimate the influence of each training sample in a target domain. Then we optimize training data of the target domain by removing the training samples that will lower the performance of the pre-trained model. We then fine-tune the pre-trained model with the optimized training data in the target domain, or build a new model which can be initialized partially based on the pre-trained model, and fine-tune it with the optimized training data in the target domain. Using this approach, transfer learning can help deep learning models to learn more useful features. Extensive experiments demonstrate the effectiveness of our approach on further boosting deep learning models for typical high-level computer vision tasks, such as image classification.
Deep learning models learn to fit training data while they are highly expected to generalize well to testing data. Most works aim at finding such models by creatively designing architectures and fine-tuning parameters. To adapt to particular tasks, hand-crafted information such as image prior has also been incorporated into end-to-end learning. However, very little progress has been made on investigating how an individual training sample will influence the generalization ability of a model. In other words, to achieve high generalization accuracy, do we really need all the samples in a training dataset? In this paper, we demonstrate that deep learning models such as convolutional neural networks may not favor all training samples, and generalization accuracy can be further improved by dropping those unfavorable samples. Specifically, the influence of removing a training sample is quantifiable, and we propose a Two-Round Training approach, aiming to achieve higher generalization accuracy. We locate unfavorable samples after the first round of training, and then retrain the model from scratch with the reduced training dataset in the second round. Since our approach is essentially different from fine-tuning or further training, the computational cost should not be a concern. Our extensive experimental results indicate that, with identical settings, the proposed approach can boost performance of the well-known networks on both high-level computer vision problems such as image classification, and low-level vision problems such as image denoising.
We observe standard transfer learning can improve prediction accuracies of target tasks at the cost of lowering their prediction fairness -- a phenomenon we named discriminatory transfer. We examine prediction fairness of a standard hypothesis transfer algorithm and a standard multi-task learning algorithm, and show they both suffer discriminatory transfer on the real-world Communities and Crime data set. The presented case study introduces an interaction between fairness and transfer learning, as an extension of existing fairness studies that focus on single task learning.
In cheminformatics, compound-target binding profiles has been a main source of data for research. For data repositories that only provide positive profiles, a popular assumption is that unreported profiles are all negative. In this paper, we caution audience not to take this assumption for granted, and present empirical evidence of its ineffectiveness from a machine learning perspective. Our examination is based on a setting where binding profiles are used as features to train predictive models; we show (1) prediction performance degrades when the assumption fails and (2) explicit recovery of unreported profiles improves prediction performance. In particular, we propose a framework that jointly recovers profiles and learns predictive model, and show it achieves further performance improvement. The presented study not only suggests applying matrix recovery methods to recover unreported profiles, but also initiates a new missing feature problem which we called Learning with Positive and Unknown Features.
In social network analysis, automatic social circle detection in ego-networks is becoming a fundamental and important task, with many potential applications such as user privacy protection or interest group recommendation. So far, most studies have focused on addressing two questions, namely, how to detect overlapping circles and how to detect circles using a combination of network structure and network node attributes. This paper asks an orthogonal research question, that is, how to detect circles based on network structures that are (usually) described by multiple views? Our investigation begins with crawling ego-networks from Twitter and employing classic techniques to model their structures by six views, including user relationships, user interactions and user content. We then apply both standard and our modified multi-view spectral clustering techniques to detect social circles in these ego-networks. Based on extensive automatic and manual experimental evaluations, we deliver two major findings: first, multi-view clustering techniques perform better than common single-view clustering techniques, which only use one view or naively integrate all views for detection, second, the standard multi-view clustering technique is less robust than our modified technique, which selectively transfers information across views based on an assumption that sparse network structures are (potentially) incomplete. In particular, the second finding makes us believe a direct application of standard clustering on potentially incomplete networks may yield biased results. We lightly examine this issue in theory, where we derive an upper bound for such bias by integrating theories of spectral clustering and matrix perturbation, and discuss how it may be affected by several network characteristics.