Data Expansion Method for Training AI Models with Limited Samples

The patent applies copying by creating synthetic copies of existing sample data through data expansion techniques. Existing samples are replicated and transformed using methods like rotation, flipping, cropping, and adding noise to generate multiple variations. This allows the system to multiply the effective training data quantity without collecting new physical samples, directly resolving the contradiction between data quantity and system complexity.

The patent employs parameter changes by systematically varying data parameters during expansion. Image parameters include rotation angles, scaling factors, brightness levels, and noise intensities. Text parameters include synonym replacement, back-translation languages, and augmentation probabilities. By changing these parameters, the system generates diverse training samples from limited originals, improving data quantity while using standardized processing pipelines that control complexity.

The patent applies preliminary action by performing data expansion and augmentation during the data preparation phase before model training begins. All synthetic data generation, including image transformations and text augmentations, is completed in advance to create an expanded training dataset. This preliminary processing allows the actual training phase to use ready-to-go data, improving fitting accuracy while separating the time cost to the offline preparation stage rather than extending training time.

The patent replaces manual data collection and annotation processes with automated data expansion algorithms. Instead of mechanically gathering and labeling new samples, the system uses computational algorithms to automatically generate synthetic variations. This substitution of mechanical data collection with algorithmic generation improves accuracy by providing more diverse training examples while reducing the time investment required for manual data preparation.