Composite Binary Decomposition Networks

Authors: You Qiaoben, Zheng Wang, Jianguo Li, Yinpeng Dong, Yu-Gang Jiang, Jun Zhu4747-4754

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate the effectiveness of the proposed method, as CBDNet can approximate image classification network ResNet-18 using 5.25 bits, VGG-16 using 5.47 bits, DenseNet-121 using 5.72 bits, object detection networks SSD300 using 4.38 bits, and semantic segmentation networks SegNet using 5.18 bits, all with minor accuracy drops.
Researcher Affiliation Collaboration You Qiaoben,1 Zheng Wang,2 Jianguo Li,3 Yinpeng Dong,1 Yu-Gang Jiang,2 Jun Zhu1 1Dept. of Comp. Sci. & Tech., State Key Lab for Intell. Tech. & Sys., Institute for AI, Tsinghua University 2School of Computer Science, Fudan University 3Intel Labs China qby222@126.com, {zhengwang17,ygj}@fudan.edu.cn, jianguo.li@intel.com, {dyp17@mails., dcszj@}tsinghua.edu.cn
Pseudocode Yes Algorithm 1 Binary matrix decomposition
Open Source Code No No explicit statement about releasing code or a direct link to a repository for the authors' implementation was found.
Open Datasets Yes The ImageNet dataset contains 1.2 million training images, 100k test images, and 50k validation images. ... The evaluation is performed on the VOC0712 dataset. ... The experiment is conducted on the Cityscapes dataset (Cordts et al. 2016).
Dataset Splits Yes The ImageNet dataset contains 1.2 million training images, 100k test images, and 50k validation images. Each image is classified into one of 1000 object categories. ... We use the validation set for evaluation and report the classification performance via top-1 and top-5 accuracies.
Hardware Specification Yes We make the inference speed comparison between CBDNet and the FP32 version on Intel Core i7-6700 CPU with 32GB RAM.
Software Dependencies No The paper mentions "TensorFlow quantizer" and "SSE4.2 instructions" but does not specify software dependencies with version numbers.
Experiment Setup Yes Figure 2 shows that J = 7 is already sufficiently good to keep a balance between the accuracy and the efficiency of the expansion. ... The effective bit-rate of CBDNet is 5.25 with b=0.3 and J=7... The effective bit-rate of CBDNet is 5.47 with b=0.3 and J=7... The effective bit-rate is 5.72 with b = 0.4 and J = 7... For CBDNet, we quantize the input into 8-bits using the Tensor Flow quantizer.