Deep clustering based on embedded auto-encoder

Deep clustering is a new research direction that combines deep learning and clustering. It performs feature representation and cluster assignments simultaneously, and its clustering performance is significantly superior to traditional clustering algorithms. The auto-encoder is a neural network model, which can learn the hidden features of the input object to achieve nonlinear dimensionality reduction. This paper proposes the embedded auto-encoder network model; specifically, the auto-encoder is embedded into the encoder unit and the decoder unit of the prototype auto-encoder, respectively. To ensure effectively cluster high-dimensional objects, the encoder of model first encodes the raw features of the input objects, and obtains a cluster-friendly feature representation. Then, in the model training stage, by adding smoothness constraints to the objective function of the encoder, the representation capabilities of the hidden layer coding are significantly improved. Finally, the adaptive self-paced learning threshold is determined according to the median distance between the object and its corresponding the centroid, and the fine-tuning sample of the model is automatically selected. Experimental results on multiple image datasets have shown that our model has fewer parameters, higher efficiency and the comprehensive clustering performance is significantly superior to the state-of-the-art clustering methods.

Automated summary

Deep clustering combines deep learning and clustering to achieve superior clustering performance. This paper proposes an embedded auto-encoder network model that effectively encodes input object features and improves clustering through smoothness constraints.