Collaborative representation induced broad learning model for classification

The broad learning system (BLS) is a novel flat neural network that is fast and effective in various pattern recognition and classification applications. Many researchers have investigated this learning approach due to its remarkable performance. However, the feature nodes used in BLS are mapped with random weights for the input data, which is inefficient and can lead to inferior results since the random mapping contains redundant and unpredictable information for constructing the feature nodes. To resolve this issue and improve BLS, in this study, we aim to present one representation induced method, i.e., the collaborative representation induced broad learning model (CRI_BLM), to replace the random mapping for producing the feature nodes. This proposed method introduces the collaborative representation technique to code the input training sample as a collaborative linear combination (coding coefficient) of all dictionary samples, before further generating the enhancement nodes under the broad learning framework for classification. Compared to the original feature nodes with random mapping, this approach can capture more effective features for pattern recognition and classification. Extensive experiments with several datasets and comparisons with various classifiers were investigated to confirm that our proposed CRI_BLM is remarkable and effective (e.g., obtaining the best result: 96.80% in the Fifteen Scene Categories database).