Can we simplify pattern recognition by using random embeddings? This research introduces Random Embedding Machines. These machines address real classification problems involving structured data. It is shown that local clustering, a set of points of a given class is linearly separable from other classes. This is achieved by embedding the points in binary space. The embedding is done by a set of randomly parameterized surfaces, with high probability. This data set is called a local relative cluster. The size of the embedding set is inversely proportional to the squared local clustering degree. This method performs as well as state-of the-art methods with much less time. By simplifying the learning problem, this research has implications for machine learning. It resolves long-standing questions about the complexity of random embedding. This approach is relevant for developing efficient pattern recognition systems.
This study on pattern recognition using Random Embedding Machines fits Neural Computation's scope, with its focus on neural networks and machine learning. The approach to simplifying classification problems and reducing computational complexity contributes to the journal's interest in efficient algorithms for neural computation.