arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 1807.09810cs.CVcs.LG
收藏

基于核集的神经网络压缩

Coreset-Based Neural Network Compression

Massachusetts Institute of Technology(麻省理工学院);University of Illinois at Urbana-Champaign(伊利诺伊大学厄巴纳-香槟分校)

Abhimanyu Dubey, Moitreya Chatterjee, Narendra Ahuja

摘要:

提出了一种基于滤波器核集表示的卷积神经网络(CNN)压缩算法。我们利用CNN权重和神经元激活(跨样本)空间中存在的冗余来获得压缩。我们的方法不需要再训练,易于实现,并在各种各样的CNN架构中获得最先进的压缩性能。再加上量化和霍夫曼编码,我们创建的网络提供了类似AlexNet的准确性,其内存占用比原始AlexNet小$832\times$,同时还显著减少了推理时间。此外,这些压缩网络在微调时,也成功地推广到其他领域。

英文摘要:

We propose a novel Convolutional Neural Network (CNN) compression algorithm based on coreset representations of filters. We exploit the redundancies extant in the space of CNN weights and neuronal activations (across samples) in order to obtain compression. Our method requires no retraining, is easy to implement, and obtains state-of-the-art compression performance across a wide variety of CNN architectures. Coupled with quantization and Huffman coding, we create networks that provide AlexNet-like accuracy, with a memory footprint that is $832\times$ smaller than the original AlexNet, while also introducing significant reductions in inference time as well. Additionally these compressed networks when fine-tuned, successfully generalize to other domains as well.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑