Over-parameterization is ubiquitous nowadays in training neural networks to benefit both optimization in seeking global optima and generalization in reducing prediction error. However, compressive networks are desired in many real world applications and direct training of small networks may be trapped in local optima. In this paper, instead of pruning or distilling over-parameterized models to compressive ones, we propose a new approach based on differential inclusions of inverse scale spaces. Specifically, it generates a family of models from simple to complex ones that couples a pair of parameters to simultaneously train over-parameterized deep models and structural sparsity on weights of fully connected and convolutional layers. Such a differential inclusion scheme has a simple discretization, proposed as Deep structurally splitting Linearized Bregman Iteration (DessiLBI), whose global convergence analysis in deep learning is established that from any initializations, algorithmic iterations converge to a critical point of empirical risks. Experimental evidence shows that DessiLBI achieve comparable and even better performance than the competitive optimizers in exploring the structural sparsity of several widely used backbones on the benchmark datasets. Remarkably, with early stopping, DessiLBI unveils "winning tickets" in early epochs: the effective sparse structure with comparable test accuracy to fully trained over-parameterized models.
We analyze the influence of adversarial training on the loss landscape of machine learning models. To this end, we first provide analytical studies of the properties of adversarial loss functions under different adversarial budgets. We then demonstrate that the adversarial loss landscape is less favorable to optimization, due to increased curvature and more scattered gradients. Our conclusions are validated by numerical analyses, which show that training under large adversarial budgets impede the escape from suboptimal random initialization, cause non-vanishing gradients and make the model find sharper minima. Based on these observations, we show that a periodic adversarial scheduling (PAS) strategy can effectively overcome these challenges, yielding better results than vanilla adversarial training while being much less sensitive to the choice of learning rate.
Graph Attention Network (GAT) and GraphSAGE are neural network architectures that operate on graph-structured data and have been widely studied for link prediction and node classification. One challenge raised by GraphSAGE is how to smartly combine neighbour features based on graph structure. GAT handles this problem through attention, however the challenge with GAT is its scalability over large and dense graphs. In this work, we proposed a new architecture to address these issues that is more efficient and is capable of incorporating different edge type information. It generates node representations by attending to neighbours sampled from weighted multi-step transition probabilities. We conduct experiments on both transductive and inductive settings. Experiments achieved comparable or better results on several graph benchmarks, including the Cora, Citeseer, Pubmed, PPI, Twitter, and YouTube datasets.
One daunting problem for semantic parsing is the scarcity of annotation. Aiming to reduce nontrivial human labor, we propose a two-stage semantic parsing framework, where the first stage utilizes an unsupervised paraphrase model to convert an unlabeled natural language utterance into the canonical utterance. The downstream naive semantic parser accepts the intermediate output and returns the target logical form. Furthermore, the entire training process is split into two phases: pre-training and cycle learning. Three tailored self-supervised tasks are introduced throughout training to activate the unsupervised paraphrase model. Experimental results on benchmarks Overnight and GeoGranno demonstrate that our framework is effective and compatible with supervised training.
Spoken Language Understanding (SLU) converts hypotheses from automatic speech recognizer (ASR) into structured semantic representations. ASR recognition errors can severely degenerate the performance of the subsequent SLU module. To address this issue, word confusion networks (WCNs) have been used to encode the input for SLU, which contain richer information than 1-best or n-best hypotheses list. To further eliminate ambiguity, the last system act of dialogue context is also utilized as additional input. In this paper, a novel BERT based SLU model (WCN-BERT SLU) is proposed to encode WCNs and the dialogue context jointly. It can integrate both structural information and ASR posterior probabilities of WCNs in the BERT architecture. Experiments on DSTC2, a benchmark of SLU, show that the proposed method is effective and can outperform previous state-of-the-art models significantly.
Few-shot learning (FSL) aims to recognize new objects with extremely limited training data for each category. Previous efforts are made by either leveraging meta-learning paradigm or novel principles in data augmentation to alleviate this extremely data-scarce problem. In contrast, this paper presents a simple statistical approach, dubbed Instance Credibility Inference (ICI) to exploit the distribution support of unlabeled instances for few-shot learning. Specifically, we first train a linear classifier with the labeled few-shot examples and use it to infer the pseudo-labels for the unlabeled data. To measure the credibility of each pseudo-labeled instance, we then propose to solve another linear regression hypothesis by increasing the sparsity of the incidental parameters and rank the pseudo-labeled instances with their sparsity degree. We select the most trustworthy pseudo-labeled instances alongside the labeled examples to re-train the linear classifier. This process is iterated until all the unlabeled samples are included in the expanded training set, i.e. the pseudo-label is converged for unlabeled data pool. Extensive experiments under two few-shot settings show that our simple approach can establish new state-of-the-arts on four widely used few-shot learning benchmark datasets including miniImageNet, tieredImageNet, CIFAR-FS, and CUB. Our code is available at: https://github.com/Yikai-Wang/ICI-FSL
Cerebral blood volume (CBV) is a hemodynamic correlate of oxygen metabolism and reflects brain activity and function. High-resolution CBV maps can be generated using the steady-state gadolinium-enhanced MRI technique. Such a technique requires an intravenous injection of exogenous gadolinium based contrast agent (GBCA) and recent studies suggest that the GBCA can accumulate in the brain after frequent use. We hypothesize that endogenous sources of contrast might exist within the most conventional and commonly acquired structural MRI, potentially obviating the need for exogenous contrast. Here, we test this hypothesis by developing and optimizing a deep learning algorithm, which we call DeepContrast, in mice. We find that DeepContrast performs equally well as exogenous GBCA in mapping CBV of the normal brain tissue and enhancing glioblastoma. Together, these studies validate our hypothesis that a deep learning approach can potentially replace the need for GBCAs in brain MRI.
The ability to extrapolate gene expression dynamics in living single cells requires robust cell segmentation, and one of the challenges is the amorphous or irregularly shaped cell boundaries. To address this issue, we modified the U-Net architecture to segment cells in fluorescence widefield microscopy images and quantitatively evaluated its performance. We also proposed a novel loss function approach that emphasizes the segmentation accuracy on cell boundaries and encourages shape feature preservation. With a 97% sensitivity, 93% specificity, 91% Jaccard similarity, and 95% Dice coefficient, our proposed method called Residual Attention U-Net with edge-enhancement surpassed the state-of-the-art U-Net in segmentation performance as evaluated by the traditional metrics. More remarkably, the same proposed candidate also performed the best in terms of the preservation of valuable shape features, namely area, eccentricity, major axis length, solidity and orientation. These improvements on shape feature preservation can serve as useful assets for downstream cell tracking and quantification of changes in cell statistics or features over time.
Certifying neural networks enables one to offer guarantees on a model's robustness. In this work, we use linear approximation to obtain an upper and lower bound of the model's output when the input data is perturbed within a predefined adversarial budget. This allows us to bound the adversary-free region in the data neighborhood by a polyhedral envelope, and calculate robustness guarantees based on this geometric approximation. Compared with existing methods, our approach gives a finer-grain quantitative evaluation of a model's robustness. Therefore, the certification method can not only obtain better certified bounds than the state-of-the-art techniques given the same adversarial budget but also derives a faster search scheme for the optimal adversarial budget. Furthermore, we introduce a simple regularization scheme based on our method that enables us to effectively train robust models.