The state-of-the-art convolutional neural network accelerators are showing a growing interest in exploiting the bit-level sparsity and eliminating the ineffectual computations of zero bits. However, the excessive redundancy and the irregular distribution of nonzero bits limit the real speedup in the accelerators. To address this, we propose an algorithm-architecture codesign, named structured term pruning (STP), to boost the computation efficiency of neural networks inference. Specifically, we enhance the bit sparsity by guiding the weights toward the value with fewer power-of-two terms. Then, we structure the terms with layer-wise group budgets. Retraining is adopted to recover the accuracy drop. We also design the hardware of the group processing element and the fast signed-digital encoder for efficient implementation of STP networks. The system design of STP is realized with some easy alterations on an input stationary systolic array design. Extensive evaluation results demonstrate that STP can reduce significant inference computation costs, and achieve$2.35\times $computational energy saving for the ResNet18 network on the ImageNet dataset.

Deep Neural Networks (DNNs) have achieved remarkable success in various Artificial Intelligence applications. Quantization is a critical step in DNNs compression and acceleration for deployment. To further boost DNN execution efficiency, many works explore to leverage the input-dependent redundancy with dynamic quantization for different regions. However, the sensitive regions in the feature map are irregularly distributed, which restricts the real speed up for existing accelerators. To this end, we propose an algorithm-architecture co-design, named Structured Dynamic Precision (SDP). Specifically, we propose a quantization scheme in which the high-order bit part and the low-order bit part of data can be masked independently. And a fixed number of term parts are dynamically selected for computation based on the importance of each term in the group. We also present a hardware design to enable the algorithm efficiently with small overheads, whose inference time mainly scales with the precision proportionally. Evaluation experiments on extensive networks demonstrate that compared to the state-of-the-art dynamic quantization accelerator DRQ, our SDP can achieve 29% performance gain and 51% energy reduction for the same level of model accuracy.

Deep neural networks have achieved remarkable advancement in various intelligence tasks. However, the massive computation and storage consumption limit applications on resource-constrained devices. While channel pruning has been widely applied to compress models, it is challenging to reach very deep compressions for such a coarse-grained pruning structure without significant performance degradation. In this article, we propose an acceleration-aware fine-grained channel pruning (AFCP) framework for accelerating neural networks, which optimizes trainable gate parameters by estimating residual errors between pruned and original channels with hardware characteristics. Our fine-grained concept consists of both algorithm and structure levels. Different from existing methods that leverage a predefined pruning criterion, AFCP explicitly considers both zero-out and similar criteria for each channel, and adaptively selects the suitable one via residual gate parameters. For structure level, AFCP adopts a fine-grained channel pruning strategy for residual neural networks and a decomposition-based structure, which further extends the pruning optimization space. Moreover, instead of using theoretical computation costs, such as floating-point operations, we propose the hardware predictor that bridges the gap between realistic acceleration and pruning procedure to guide the learning of pruning, which improves the efficiency of model pruning when deployed on accelerators. Extensive evaluation results demonstrate that AFCP outperforms state-of-the-art methods, and achieves a favorable balance between model performance and computation cost.

With the development of deployment platforms and application scenarios for deep neural networks, traditional fixed network architectures cannot meet the requirements. Meanwhile the dynamic network inference becomes a new research trend. Many slimmable and scalable networks have been proposed to satisfy different resource constraints (e.g., storage, latency and energy). And a single network may support versatile architectural configurations including: depth, width, kernel size, and resolution. In this paper, we propose a novel network architecture reuse strategy enabling dynamic precision in parameters. Since our low-precision networks are wrapped in the high-precision networks like an onion, we name it dynamic precision onion quantization (DPOQ). We train the network by using the joint loss with scaled gradients. To further improve the performance and make different precision network compatible with each other, we propose the precision shift batch normalization (PSBN). And we also propose a scalable input-specific inference mechanism based on this architecture and make the network more adaptable. Experiments on the CIFAR and ImageNet dataset have shown that our DPOQ achieves not only better flexibility but also higher accuracy than the individual quantization.

Data free deep neural network quantization is a practical challenge, since the original training data is often unavailable due to some privacy, proprietary or transmission issues. The existing methods implicitly equate data-free with training-free and quantize model manually through analyzing the weights’ distribution. It leads to a significant accuracy drop in lower than 6-bit quantization. In this work, we propose the data free quantization-aware fine-tuning (DFQF), wherein no real training data is required, and the quantized network is fine-tuned with generated images. Specifically, we start with training a generator from the pre-trained full-precision network with inception score loss, batch-normalization statistics loss and adversarial loss to synthesize a fake image set. Then we fine-tune the quantized student network with the full-precision teacher network and the generated images by utilizing knowledge distillation (KD). The proposed DFQF outperforms state-of-the-art post-train quantization methods, and achieve W4A4 quantization of ResNet20 on the CIFAR10 dataset within 1% accuracy drop.