CNN-Based Broad Learning System
Ting Li, Bin Fang, Jiye Qian, Xuegang Wu · 2019
The Broad Learning System (BLS) has recently been proposed as an effective and useful method for pattern recognition. It overcomes the shortcomings of deep neural networks that are both time consuming and hardware dependent. BLS is a lightweight network structure. Although its advantages in speed and flexibility are obvious, there is a still a gap between its accuracy in image recognition and that of state-of-art approaches. We propose a BLS-based model that combines CNN and Adam algorithm. At first, we build a BLS-based classifier, and then we build a CNN-based feature extraction. After that, we take features extracted as inputs to the classifier. Feature extraction through convolution and pooling operations can effectively preserve key features of the image. We use Adam algorithm to update the weights of the feature extraction. By retaining key features and removing factors that affect image recognition, it helps to improve the accuracy of the classification results. Therefore, through the above operations, the speed and accuracy of our proposed lightweight network structure haves been improved and reflected in the dataset of the MNIST and Cat or Dog for Kaggle competitions. We can choose different convolution layers, pooling layers and optimization algorithms according to the different sized data sets.