Topological mappings of video and audio data

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Visualization

Background:

• Self-organizing maps (SOMs) are unsupervised learning algorithms used for data visualization and clustering.
• Generative Topographic Mapping (GTM) is a probabilistic model extending mixture of experts.
• Existing methods may not fully capture complex data structures.

Purpose of the Study:

• To introduce and evaluate a new self-organizing map based on a product of experts model.
• To compare the performance of the new algorithm against the standard Self-Organizing Map and GTM.
• To demonstrate the algorithm's effectiveness on video data of Korean vowel utterances.

Main Methods:

• The proposed model utilizes a nonlinear projection of latent points, similar to GTM, but extends a product of experts framework.
• The algorithm was tested using a dataset of video recordings of lips articulating five Korean vowels.
• An alternative algorithm was derived by minimizing mean squared error, incorporating both local and global information.

Main Results:

• The new self-organizing map demonstrated superior visualization and clustering results compared to the standard SOM.
• The product of experts model showed effectiveness in analyzing complex, high-dimensional data like speech.
• The mean squared error minimization approach also yielded improved clustering performance.

Conclusions:

• The developed self-organizing map, an extension of the product of experts, provides enhanced performance for data clustering and visualization.
• The algorithm's ability to integrate local and global information offers a significant advantage over traditional methods.
• This research contributes a more effective unsupervised learning technique for complex pattern recognition tasks.