KM-DBSCAN: an enhanced density and centroid based border detection framework for data reduction towards green AI.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• The exponential growth of training data increases computational costs and energy consumption in AI.
• Existing AI model compression techniques include pruning, quantization, and knowledge distillation.
• Data reduction is a key strategy for enhancing training speed and the 'green AI' score.

Purpose of the Study:

• Introduce KM-DBSCAN, a new data clustering algorithm for intelligent data reduction.
• Combine K-Means and DBSCAN properties for efficient data clustering and border detection.
• Evaluate the impact of data reduction on machine learning model performance and environmental efficiency.

Main Methods:

• Developed KM-DBSCAN, a hybrid clustering algorithm merging K-Means and DBSCAN.
• Applied KM-DBSCAN for data reduction on six benchmark datasets (Banana, USPS, Adult9a, Collision, Dry Bean, Melanoma).
• Trained and tested Support Vector Machines (SVM), Multilayer Perceptrons (MLP), and Convolutional Neural Networks (CNN) on reduced datasets.

Main Results:

• Achieved up to 90% data reduction with KM-DBSCAN.
• Observed training speedups ranging from 3.6× to 6900×.
• Demonstrated significant reductions in carbon emissions (0.0219 g to 5.374 g) while maintaining competitive accuracy, e.g., 90.39% for melanoma classification with 28.7% data and 71.65% emission reduction.

Conclusions:

• KM-DBSCAN enables substantial data reduction, leading to faster training and lower energy consumption.
• The algorithm effectively reduces carbon footprint in AI model training.
• KM-DBSCAN facilitates environmentally conscious AI development without sacrificing predictive performance.