Kohonen Network-Based Adaptation of Non Sequential Data for Use in Convolutional Neural Networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Convolutional Neural Networks (CNNs) excel at analyzing sequential data like images and time series.
• CNNs are not inherently suited for non-sequential feature vector data where element order is irrelevant.
• Existing methods lack effective ways to apply CNNs to unordered data.

Purpose of the Study:

• To develop a novel method for transforming non-sequential data into an image format compatible with CNN analysis.
• To enhance the applicability of powerful CNN models to a broader range of data types.
• To evaluate the effectiveness of the proposed method on benchmark datasets and a challenging real-world problem.

Main Methods:

• Utilized Kohonen networks (self-organizing maps) to train on non-sequential data.
• Developed a 'U-image' representation for each data example, converting it into a format suitable for CNN input.
• Implemented a hybrid approach combining U-image representation with original features for a dual input signal.

Main Results:

• Successfully represented non-sequential data as U-images for CNN processing.
• Demonstrated the method's efficacy on traditional machine learning databases.
• Achieved positive results on a complex classification task using experimental particle physics data.

Conclusions:

• The proposed U-image representation method effectively adapts non-sequential data for CNN analysis.
• This technique broadens the scope of CNN applications beyond traditional sequential data.
• The hybrid approach shows promise for improving classification performance in challenging domains.