Deep Learning Model Compression for Image Analysis: Methods and Architectures

网址链接：https://medium.com/zylapp/deep-learning-model-compression-for-image-analysis-methods-and-architectures-398f82b0c06f

PDF链接：https://pan.baidu.com/s/1HLHL20Codfshe9jbhXaQIQ；提取码：i7mh

In my previous blog posts, I have reviewed well-known and recent deep learning algorithms for image classification and object detection. At Zyl, we aim to embed deep learning models directly within user’s smartphones to guarantee their privacy. This induces multiple constraints for these models known to be heavy and greedy in energy. Fortunately, at Zyl the embedded models are only used for inference1. Nevertheless, the computational cost is high even for inference and requires a lot of energy.

This blog post describes theoretical methods to reduce model size. Size reduction for deep learning models is an active field of research. Those methods are truly performant, but the specific type of machine learning models used involves extremely deep and complex architectures (Simonyan et al. (2014), He et al. (2015), Szegedy et al. (2016)). How can we simply transform a deep model into a lighter one without decreasing drastically its performances ? Moreover, does it exist specialized architectures to build light models while achieving state-of-the-art performances ? Note that researchers test their algorithms using different datasets. Thus the cited accuracies cannot be directly compared per se.