Recent years have witnessed growing interests in designing efficient neural networks and neural architecture search (NAS). Although remarkable efficiency and accuracy have been achieved, existing expert designed and NAS models neglect the fact that input instances are of varying complexity thus different amount of computation is required. Inference with a fixed model that processes all instances through the same transformations would waste plenty of computational resources. Therefore, customizing the model capacity in an instance-aware manner is highly demanded. To address this issue, we propose an Instance-aware Selective Branching Network-ISBNet, which supports efficient instance-level inference by selectively bypassing transformation branches of insignificant importance weight. These weights are determined dynamically by accompanying lightweight hypernetworks SelectionNets and further recalibrated by gumbel-softmax for sparse branch selection. Extensive experiments show that ISBNet achieves extremely efficient inference in terms of parameter size and FLOPs comparing to existing networks. For example, ISBNet takes only 8.03% parameters and 30.60% FLOPs of the state-of-the-art efficient network ShuffleNetV2 with comparable accuracy.
Machine-learning-based data-driven applications have become ubiquitous, e.g., health-care analysis and database system optimization. Big training data and large (deep) models are crucial for good performance. Dropout has been widely used as an efficient regularization technique to prevent large models from overfitting. However, many recent works show that dropout does not bring much performance improvement for deep convolutional neural networks (CNNs), a popular deep learning model for data-driven applications. In this paper, we formulate existing dropout methods for CNNs under the same analysis framework to investigate the failures. We attribute the failure to the conflicts between the dropout and the batch normalization operation after it. Consequently, we propose to change the order of the operations, which results in new building blocks of CNNs.Extensive experiments on benchmark datasets CIFAR, SVHN and ImageNet have been conducted to compare the existing building blocks and our new building blocks with different dropout methods. The results confirm the superiority of our proposed building blocks due to the regularization and implicit model ensemble effect of dropout. In particular, we improve over state-of-the-art CNNs with significantly better performance of 3.17%, 16.15%, 1.44%, 21.46% error rate on CIFAR-10, CIFAR-100, SVHN and ImageNet respectively.
Deep learning models have been used to support analytics beyond simple aggregation, where deeper and wider models have been shown to yield great results. These models consume a huge amount of memory and computational operations. However, most of the large-scale industrial applications are often computational budget constrained. Current solutions are mainly based on model compression -- deploying a smaller model to save the computational resources. Meanwhile, the peak workload of inference service could be 10x higher than the average cases, with even unpredictable extreme cases. Lots of computational resources could be wasted during off-peak hours. On the other hand, the system may crash when the workload exceeds system design. Supporting such deep learning service with dynamic workload cost-efficiently remains to be a challenging problem. We address this conflict with a general and novel training scheme called model slicing, which enables deep learning models to provide predictions within prescribed computational resource budget dynamically. Model slicing could be viewed as an elastic computation solution without requiring more computation resources, but by slightly sacrificing prediction accuracy. In a nutshell, partially ordered relation is introduced to the basic components of each layer in the model, namely neurons in dense layers and channels in convolutional layers. Specifically, if one component participates in the forward pass, then all of its preceding components are also activated. Dynamically trained under such structural constraint, we can slice a narrower sub-model during inference whose run-time memory and computational operation consumption is roughly quadratic to the width controlled by a single parameter slice rate. Extensive experiments show that models trained with model slicing can support elastic inference cost effectively with minimum performance loss.