Layer-Wise Compression Analysis of CNN Architectures: Insights from VGG16

The state-of-the-art performance of most visual tasks is accomplished by deep convolutional neural networks, which have millions to billions of parameters, resulting in large neural networks. This introduced more redundant and oversized layers, which pose challenges in terms of memory requirements and computational efficiency. To address these issues, this paper investigates the compressibility of CNN layers, with the aim of reducing the size of the model. In this paper, we propose a method based on an autoencoder, specifically CSNet (Compressive Sensing Network) for layer-wise compression analysis. Our study reveals that CNN models can be effectively compressed using autoencoder-based techniques. We assess the performance drop by compressing individual layers and combinations thereof, providing insight into the layers that contribute the most to the inference result. Our contributions include training and evaluating an autoencoder compression network (CSNet), to compress layer-per-layer a CNN model (VGG16), in order to conduct a comprehensive analysis of CNN compressibility and evaluating the performance impact of layer-wise compression.