Figure 1.
CNN Architecture for CIFAR-10
Table 1.
Comprehensive performance analysis across augmentation strategies
| Method (Intensity Score) | Val Acc (%) | F1-Score | Training Time (s) | Overfitting gap (%) |
|---|---|---|---|---|
| No Augmentation (0.0) | 77.49 | 0.774 | 650.4 | 3.54 |
| Basic (0.49) | 79.84 | 0.797 | 1255.6 | -1.56 |
| Light Advanced (0.09) | 78.80 | 0.786 | 342.5 | -0.28 |
| Moderate Advanced (0.51) | 75.59 | 0.754 | 341.7 | -4.77 |
| Strong Advanced (0.94) | 71.64 | 0.714 | 343.5 | -13.06 |
| AutoAugment Style (0.98) | 74.01 | 0.737 | 342.2 | -6.83 |
Figure 2.
Performance vs. Augmentation Intensity
Figure 3.
Comprehensive performance comparison across all augmentation methods
Figure 4.
Learning curve comparison
Figure 5.
Overfitting Gap Analysis
Table 2.
Intensity Framework Validation Through Component Analysis
| Intensity Range & Strategies | Performance | Transformation Diversity | Parameter Impact |
|---|---|---|---|
| Baseline (0.0): No Augmentation | 77.49% | 0 transformations Only preprocessing: Resize to 224×224, ImageNet normalization No regularization benefit | Maintains original data fidelity but lacks regularization capacity, leading to overfitting on training data |
| Light (0.09) | 78.80% | 2 transformations Conservative diversity: HorizontalFlip (p=0.5), Random Brightness Contrast (p=0.3) Minimal but effective regularization | Moderate parameters balance regularization and stability |
| Optimal (0.49): Basic | 79.84% | 3 transformations Optimal diversity balance: RandomHorizontalFlip (p=0.5), RandomRotation (±10°), ColorJitter (brightness, contrast, saturation ±0.2)Perfect regularization-performance trade-off | Moderate parameters Rotation ±10°, color jitter ±0.2 range achieves optimal balance between regularization effectiveness and learning stability |
| Moderate (0.51): Moderate Advanced | 75.59% | 4 transformations Increased complexity: HorizontalFlip (p=0.5), ShiftScaleRotate (p=0.4), Random Brightness Contrast (p=0.4), Hue Saturation Value (p=0.3) Complexity begins to create interference | Aggressive parameters Shift/scale ±0.1, rotation ±15°, HSV modifications create increased parameter ranges that start introducing instability |
| Heavy (0.94-0.98): Strong Advanced, AutoAugment Style | 71.64%-74.01% | 5-6 transformations Excessive complexity: Multiple geometric transforms, destructive elements (CoarseDropout, GaussNoise), Complex photometric (GridDistortion, RandomGamma) Overwhelming learning capacity | Aggressive parameters Rotation ±25°, noise injection, aggressive parameter ranges (±0.2+) distort data distribution beyond model’s learning capacity |