A Multiobjective Sparse Feature Learning Model for Deep Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Hierarchical deep neural networks mimic human brain architecture.
• Sparse feature learning is crucial for effective representation, but often requires manual sparsity control.
• Existing models necessitate user-defined constants for sparsity regulation.

Purpose of the Study:

• To propose a multiobjective sparse feature learning model based on autoencoders.
• To automatically find a compromise between reconstruction error and feature sparsity.
• To develop an effective learning procedure for the proposed model.

Main Methods:

• Developed a multiobjective sparse feature learning model utilizing an autoencoder architecture.
• Optimized model parameters by simultaneously considering reconstruction error and hidden unit sparsity.
• Designed a multiobjective induced learning procedure employing a multiobjective evolutionary algorithm.

Main Results:

• Demonstrated the effectiveness of the proposed learning procedure through experimental evaluation.
• Showcased the model's capability to learn useful sparse features.
• Validated the automatic balancing of reconstruction error and sparsity.

Conclusions:

• The proposed multiobjective sparse feature learning model effectively learns valuable sparse representations.
• The developed learning procedure is efficient and removes the need for manual sparsity constant definition.
• This approach offers an automated method for sparse feature learning in deep networks.