Adaptive Template Reconstruction for Effective Pattern Classification.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Pattern classification often struggles with limited or noisy training data.
• Existing algorithms may not perform optimally under such challenging conditions.

Purpose of the Study:

• To introduce and evaluate a novel instance-based algorithm for pattern classification.
• To address the limitations of existing methods when dealing with scarce and noisy datasets.

Main Methods:

• The proposed algorithm transforms query data and training templates based on feature space distributions.
• A key novelty is template reconstruction, enhancing performance with limited training data.
• The method was evaluated on image (FASHION-MNIST, CIFAR-10) and time-series (EEG) datasets.

Main Results:

• Achieved a 2-3% average classification improvement on image datasets using small training subsets compared to state-of-the-art methods on full datasets.
• Demonstrated effectiveness in classifying non-stationary, noisy electroencephalography (EEG) signals.
• Adaptive reconstruction of feature instances significantly improved class separation and matching for both images and time-series.

Conclusions:

• The novel algorithm shows significant improvements in pattern classification, especially with limited and noisy data.
• Its versatility is proven across image and time-series data, including challenging EEG signals.
• The method holds potential for diverse applications requiring robust classification under data constraints.