Month: September 2026

Early CNN layers learn universal visual features that transfer across tasks. Only later layers learn task-specific patterns. Therefore: borrow the universal early layers from a large pre-trained model, discard the final task-specific layers, and train only the new head on your small dataset.