本文为荷兰拉德堡德大学（作者：Erdi C¸ allı）的硕士论文，共58页。

最先进的图像处理模型的计算代价与公开可用的设备处理能力之间存在着间隙，这一差距正在降低这些有希望模型的适用性。为了弥补这一差距，我们首先研究了裁剪和因子分解以降低模型的计算成本。其次，我们寻找可供选择的卷积运算来设计最先进的模型。第三，利用这些可替代的卷积运算，我们训练了CIFAR-10分类任务的模型。我们提出的模型在一半的模型大小和三分之一的浮点运算下，获得了与ResNet-20（91:25%的top-1精度）相当的结果（91:1%的top-1精度）。最后，我们使用了裁剪和因子分解的设计，发现这些方法在降低计算复杂度和保持我们所提出模型的准确性方面是无效的。

There exists a gap between the computational cost of state of theart image processing models and the processing power of publicly availabledevices. This gap is reducing the applicability of these promising models.Trying to bridge this gap, first we investigate pruning and factorization toreduce the computational cost of a model. Secondly, we look for alternativeconvolution operations to design state of the art models. Thirdly, using thesealternative convolution operations, we train a model for the CIFAR-10classification task. Our proposed model achieves comparable results (91:1%top-1 accuracy) to ResNet-20 (91:25% top-1accuracy) with half the model size and one-third floating point operations.Finally, we apply pruning and factorization and observe that these methods areineffective in reducing the computational complexity and preserving theaccuracy of our proposed model.

1. 引言

2. 研究方法

3. 结果

4. 讨论

5. 结论

更多精彩文章请关注公众号：