Deepening Neural Networks Implicitly and Locally via Recurrent Attention Strategy

More and more empirical and theoretical evidence shows that deepening neural networks can effectively improve their performance under suitable training settings. However, deepening the backbone of neural networks will inevitably and significantly increase computation and parameter size. To mitigate these problems, we propose a simple-yet-effective Recurrent Attention Strategy (RAS), which implicitly increases the depth of neural networks with lightweight attention modules by local parameter sharing. The extensive experiments on three widely-used benchmark datasets demonstrate that RAS can improve the performance of neural networks at a slight addition of parameter size and computation, performing favorably against other existing well-known attention modules.

Causal Inference for Chatting Handoff

Aiming to ensure chatbot quality by predicting chatbot failure and enabling human-agent collaboration, Machine-Human Chatting Handoff (MHCH) has attracted lots of attention from both industry and academia in recent years. However, most existing methods mainly focus on the dialogue context or assist with global satisfaction prediction based on multi-task learning, which ignore the grounded relationships among the causal variables, like the user state and labor cost. These variables are significantly associated with handoff decisions, resulting in prediction bias and cost increasement. Therefore, we propose Causal-Enhance Module (CEM) by establishing the causal graph of MHCH based on these two variables, which is a simple yet effective module and can be easy to plug into the existing MHCH methods. For the impact of users, we use the user state to correct the prediction bias according to the causal relationship of multi-task. For the labor cost, we train an auxiliary cost simulator to calculate unbiased labor cost through counterfactual learning so that a model becomes cost-aware. Extensive experiments conducted on four real-world benchmarks demonstrate the effectiveness of CEM in generally improving the performance of existing MHCH methods without any elaborated model crafting.

Accelerating Numerical Solvers for Large-Scale Simulation of Dynamical System via NeurVec

Ensemble-based large-scale simulation of dynamical systems is essential to a wide range of science and engineering problems. Conventional numerical solvers used in the simulation are significantly limited by the step size for time integration, which hampers efficiency and feasibility especially when high accuracy is desired. To overcome this limitation, we propose a data-driven corrector method that allows using large step sizes while compensating for the integration error for high accuracy. This corrector is represented in the form of a vector-valued function and is modeled by a neural network to regress the error in the phase space. Hence we name the corrector neural vector (NeurVec). We show that NeurVec can achieve the same accuracy as traditional solvers with much larger step sizes. We empirically demonstrate that NeurVec can accelerate a variety of numerical solvers significantly and overcome the stability restriction of these solvers. Our results on benchmark problems, ranging from high-dimensional problems to chaotic systems, suggest that NeurVec is capable of capturing the leading error term and maintaining the statistics of ensemble forecasts.

The Lottery Ticket Hypothesis for Self-attention in Convolutional Neural Network

Recently many plug-and-play self-attention modules (SAMs) are proposed to enhance the model generalization by exploiting the internal information of deep convolutional neural networks (CNNs). In general, previous works ignore where to plug in the SAMs since they connect the SAMs individually with each block of the entire CNN backbone for granted, leading to incremental computational cost and the number of parameters with the growth of network depth. However, we empirically find and verify some counterintuitive phenomena that: (a) Connecting the SAMs to all the blocks may not always bring the largest performance boost, and connecting to partial blocks would be even better; (b) Adding the SAMs to a CNN may not always bring a performance boost, and instead it may even harm the performance of the original CNN backbone. Therefore, we articulate and demonstrate the Lottery Ticket Hypothesis for Self-attention Networks: a full self-attention network contains a subnetwork with sparse self-attention connections that can (1) accelerate inference, (2) reduce extra parameter increment, and (3) maintain accuracy. In addition to the empirical evidence, this hypothesis is also supported by our theoretical evidence. Furthermore, we propose a simple yet effective reinforcement-learning-based method to search the ticket, i.e., the connection scheme that satisfies the three above-mentioned conditions. Extensive experiments on widely-used benchmark datasets and popular self-attention networks show the effectiveness of our method. Besides, our experiments illustrate that our searched ticket has the capacity of transferring to some vision tasks, e.g., crowd counting and segmentation.

AlterSGD: Finding Flat Minima for Continual Learning by Alternative Training

Deep neural networks suffer from catastrophic forgetting when learning multiple knowledge sequentially, and a growing number of approaches have been proposed to mitigate this problem. Some of these methods achieved considerable performance by associating the flat local minima with forgetting mitigation in continual learning. However, they inevitably need (1) tedious hyperparameters tuning, and (2) additional computational cost. To alleviate these problems, in this paper, we propose a simple yet effective optimization method, called AlterSGD, to search for a flat minima in the loss landscape. In AlterSGD, we conduct gradient descent and ascent alternatively when the network tends to converge at each session of learning new knowledge. Moreover, we theoretically prove that such a strategy can encourage the optimization to converge to a flat minima. We verify AlterSGD on continual learning benchmark for semantic segmentation and the empirical results show that we can significantly mitigate the forgetting and outperform the state-of-the-art methods with a large margin under challenging continual learning protocols.

Blending Pruning Criteria for Convolutional Neural Networks

The advancement of convolutional neural networks (CNNs) on various vision applications has attracted lots of attention. Yet the majority of CNNs are unable to satisfy the strict requirement for real-world deployment. To overcome this, the recent popular network pruning is an effective method to reduce the redundancy of the models. However, the ranking of filters according to their "importance" on different pruning criteria may be inconsistent. One filter could be important according to a certain criterion, while it is unnecessary according to another one, which indicates that each criterion is only a partial view of the comprehensive "importance". From this motivation, we propose a novel framework to integrate the existing filter pruning criteria by exploring the criteria diversity. The proposed framework contains two stages: Criteria Clustering and Filters Importance Calibration. First, we condense the pruning criteria via layerwise clustering based on the rank of "importance" score. Second, within each cluster, we propose a calibration factor to adjust their significance for each selected blending candidates and search for the optimal blending criterion via Evolutionary Algorithm. Quantitative results on the CIFAR-100 and ImageNet benchmarks show that our framework outperforms the state-of-the-art baselines, regrading to the compact model performance after pruning.

Continuous Transition: Improving Sample Efficiency for Continuous Control Problems via MixUp

Although deep reinforcement learning~(RL) has been successfully applied to a variety of robotic control tasks, it's still challenging to apply it to real-world tasks, due to the poor sample efficiency. Attempting to overcome this shortcoming, several works focus on reusing the collected trajectory data during the training by decomposing them into a set of policy-irrelevant discrete transitions. However, their improvements are somewhat marginal since i) the amount of the transitions is usually small, and ii) the value assignment only happens in the joint states. To address these issues, this paper introduces a concise yet powerful method to construct \textit{Continuous Transition}, which exploits the trajectory information by exploiting the potential transitions along the trajectory. Specifically, we propose to synthesize new transitions for training by linearly interpolating the conjunctive transitions. To keep the constructed transitions authentic, we also develop a discriminator to guide the construction process automatically. Extensive experiments demonstrate that our proposed method achieves a significant improvement in sample efficiency on various complex continuous robotic control problems in MuJoCo and outperforms the advanced model-based / model-free RL methods.

Efficient Attention Network: Accelerate Attention by Searching Where to Plug

Recently, many plug-and-play self-attention modules are proposed to enhance the model generalization by exploiting the internal information of deep convolutional neural networks (CNNs). Previous works lay an emphasis on the design of attention module for specific functionality, e.g., light-weighted or task-oriented attention. However, they ignore the importance of where to plug in the attention module since they connect the modules individually with each block of the entire CNN backbone for granted, leading to incremental computational cost and number of parameters with the growth of network depth. Thus, we propose a framework called Efficient Attention Network (EAN) to improve the efficiency for the existing attention modules. In EAN, we leverage the sharing mechanism (Huang et al. 2020) to share the attention module within the backbone and search where to connect the shared attention module via reinforcement learning. Finally, we obtain the attention network with sparse connections between the backbone and modules, while (1) maintaining accuracy (2) reducing extra parameter increment and (3) accelerating inference. Extensive experiments on widely-used benchmarks and popular attention networks show the effectiveness of EAN. Furthermore, we empirically illustrate that our EAN has the capacity of transferring to other tasks and capturing the informative features. The code is available at https://github.com/gbup-group/EAN-efficient-attention-network

Convolution-Weight-Distribution Assumption: Rethinking the Criteria of Channel Pruning

Channel pruning is one of the most important techniques for compressing neural networks with convolutional filters. However, in our study, we find strong similarities among some primary pruning criteria proposed in recent years. The sequence of filters'"importance" in a convolutional layer according to these criteria are almost the same, resulting in similar pruned structures. This finding can be explained by our assumption that the trained convolutional filters approximately follow a Gaussian-alike distribution, which is demonstrated through systematic and comprehensive statistical tests. Under this assumption, the similarity of these criteria is theoretically proved. Moreover, we also find that if the network has too much redundancy(exists a large number of filters in each convolutional layer), then these criteria can not distinguish the "importance" of the filters. This phenomenon is due to that the convolutional layer will form a special geometric structure when redundancy is large enough and our assumption holds: for every pair of filters in one layer, (1)Their $\ell_2$ norm are equivalent; (2)They are equidistant; (3)and they are orthogonal. The full appendix is released at https://github.com/dedekinds/CWDA.

Instance Enhancement Batch Normalization: an Adaptive Regulator of Batch Noise

Batch Normalization (BN) (Ioffe and Szegedy 2015) normalizes the features of an input image via statistics of a batch of images and this batch information is considered as batch noise that will be brought to the features of an instance by BN. We offer a point of view that self-attention mechanism can help regulate the batch noise by enhancing instance-specific information. Based on this view, we propose combining BN with a self-attention mechanism to adjust the batch noise and give an attention-based version of BN called Instance Enhancement Batch Normalization (IEBN) which recalibrates channel information by a simple linear transformation. IEBN outperforms BN with a light parameter increment in various visual tasks universally for different network structures and benchmark data sets. Besides, even if under the attack of synthetic noise, IEBN can still stabilize network training with good generalization. The code of IEBN is available at https://github.com/gbup-group/IEBN