Qing Yang

Sheng Yang, Tong Zhan, Shichen Qiao et al.

Reliable 3D object perception is essential in autonomous driving. Owing to its sensing capabilities in all weather conditions, 4D radar has recently received much attention. However, compared to LiDAR, 4D radar provides much sparser point cloud. In this paper, we propose a 3D object detection method, termed ZFusion, which fuses 4D radar and vision modality. As the core of ZFusion, our proposed FP-DDCA (Feature Pyramid-Double Deformable Cross Attention) fuser complements the (sparse) radar information and (dense) vision information, effectively. Specifically, with a feature-pyramid structure, the FP-DDCA fuser packs Transformer blocks to interactively fuse multi-modal features at different scales, thus enhancing perception accuracy. In addition, we utilize the Depth-Context-Split view transformation module due to the physical properties of 4D radar. Considering that 4D radar has a much lower cost than LiDAR, ZFusion is an attractive alternative to LiDAR-based methods. In typical traffic scenarios like the VoD (View-of-Delft) dataset, experiments show that with reasonable inference speed, ZFusion achieved the state-of-the-art mAP (mean average precision) in the region of interest, while having competitive mAP in the entire area compared to the baseline methods, which demonstrates performance close to LiDAR and greatly outperforms those camera-only methods.

Xu Ma, Jingda Guo, Sihai Tang et al.

While self-attention mechanism has shown promising results for many vision tasks, it only considers the current features at a time. We show that such a manner cannot take full advantage of the attention mechanism. In this paper, we present Deep Connected Attention Network (DCANet), a novel design that boosts attention modules in a CNN model without any modification of the internal structure. To achieve this, we interconnect adjacent attention blocks, making information flow among attention blocks possible. With DCANet, all attention blocks in a CNN model are trained jointly, which improves the ability of attention learning. Our DCANet is generic. It is not limited to a specific attention module or base network architecture. Experimental results on ImageNet and MS COCO benchmarks show that DCANet consistently outperforms the state-of-the-art attention modules with a minimal additional computational overhead in all test cases. All code and models are made publicly available.

Yun Ye, Ganmei You, Jong-Kae Fwu et al.

Structured pruning, especially channel pruning is widely used for the reduced computational cost and the compatibility with off-the-shelf hardware devices. Among existing works, weights are typically removed using a predefined global threshold, or a threshold computed from a predefined metric. The predefined global threshold based designs ignore the variation among different layers and weights distribution, therefore, they may often result in sub-optimal performance caused by over-pruning or under-pruning. In this paper, we present a simple yet effective method, termed Optimal Thresholding (OT), to prune channels with layer dependent thresholds that optimally separate important from negligible channels. By using OT, most negligible or unimportant channels are pruned to achieve high sparsity while minimizing performance degradation. Since most important weights are preserved, the pruned model can be further fine-tuned and quickly converge with very few iterations. Our method demonstrates superior performance, especially when compared to the state-of-the-art designs at high levels of sparsity. On CIFAR-100, a pruned and fine-tuned DenseNet-121 by using OT achieves 75.99% accuracy with only 1.46e8 FLOPs and 0.71M parameters.

Qing Yang, Jiachen Mao, Zuoguan Wang et al.

To improve the execution speed and efficiency of neural networks in embedded systems, it is crucial to decrease the model size and computational complexity. In addition to conventional compression techniques, e.g., weight pruning and quantization, removing unimportant activations can reduce the amount of data communication and the computation cost. Unlike weight parameters, the pattern of activations is directly related to input data and thereby changes dynamically. To regulate the dynamic activation sparsity (DAS), in this work, we propose a generic low-cost approach based on winners-take-all (WTA) dropout technique. The network enhanced by the proposed WTA dropout, namely \textit{DASNet}, features structured activation sparsity with an improved sparsity level. Compared to the static feature map pruning methods, DASNets provide better computation cost reduction. The WTA technique can be easily applied in deep neural networks without incurring additional training variables. More importantly, DASNet can be seamlessly integrated with other compression techniques, such as weight pruning and quantization, without compromising on accuracy. Our experiments on various networks and datasets present significant run-time speedups with negligible accuracy loss.

Qing Yang, Wei Wen, Zuoguan Wang et al.

With the rapid scaling up of deep neural networks (DNNs), extensive research studies on network model compression such as weight pruning have been performed for improving deployment efficiency. This work aims to advance the compression beyond the weights to neuron activations. We propose the joint regularization technique which simultaneously regulates the distribution of weights and activations. By distinguishing and leveraging the significance difference among neuron responses and connections during learning, the jointly pruned network, namely \textit{JPnet}, optimizes the sparsity of activations and weights for improving execution efficiency. The derived deep sparsification of JPnet reveals more optimization space for the existing DNN accelerators dedicated for sparse matrix operations. We thoroughly evaluate the effectiveness of joint regularization through various network models with different activation functions and on different datasets. With $0.4\%$ degradation constraint on inference accuracy, a JPnet can save $72.3\% \sim 98.8\%$ of computation cost compared to the original dense models, with up to $5.2\times$ and $12.3\times$ reductions in activation and weight numbers, respectively.