☆ 4.6 Article

Gradient Estimation for Ultra Low Precision POT and Additive POT Quantization

IEEE ACCESS (2023)

Distribution-Sensitive Information Retention for Accurate Binary Neural Network

Haotong Qin et al.

Summary: Model binarization is an effective method for compressing neural networks. This paper introduces a novel distribution-sensitive information retention network (DIR-Net) that improves the performance of binary neural networks (BNNs). DIR-Net utilizes three technical contributions to retain information in the forward and backward propagation processes. Comprehensive experiments show that DIR-Net outperforms state-of-the-art binarization approaches and achieves storage saving and speedup on resource-limited devices.

INTERNATIONAL JOURNAL OF COMPUTER VISION (2023)

添加到收藏夹

Article Computer Science, Artificial Intelligence

Diverse Sample Generation: Pushing the Limit of Generative Data-Free Quantization

Haotong Qin et al.

Summary: Generative data-free quantization is a compression approach that quantizes deep neural networks to low bit-width without accessing the real data. This paper presents a generic Diverse Sample Generation (DSG) scheme to mitigate the accuracy degradation issue through generating diverse synthetic samples.

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE (2023)

添加到收藏夹

Article Engineering, Electrical & Electronic

Quantization-Error-Robust Deep Neural Network for Embedded Accelerators

Youngbeom Jung et al.

Summary: Quantization with low precision is an important technique for adopting deep neural networks in energy- and memory-constrained devices. This study proposes methods for reforming and quantizing a network to achieve high accuracy even at low precision without any runtime overhead in embedded accelerators. Experimental results show accuracies of up to 98.31% and 99.96% in 6-bit and 8-bit quantization networks, respectively.

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II-EXPRESS BRIEFS (2022)

添加到收藏夹

Article Engineering, Electrical & Electronic