What fine-tuning changes AI
Fine-tuning continues training a model that already learned from another dataset. The image pilot begins with ResNet-18 weights learned on ImageNet, then replaces its original classifier with a randomly initialized 37-breed classifier.
For the first epoch, only that new classifier learns. The backbone still processes each image, but its parameters are frozen. The next two epochs update both backbone and classifier. Thus the same crop can acquire a different representation as well as a different predicted breed. PyTorch’s transfer-learning tutorial distinguishes training a new classifier from adapting the pretrained network.
At the transition, this pilot also resets the optimizer and changes learning rates. Validation accuracy falls from 85.3% to 72.8% at the next saved checkpoint, then reaches 86.9%. Changing features are a plausible explanation for disruption, but the run does not isolate the contributions of unfreezing, optimizer reset, or learning rate.
The activity GRU is trained from random initialization. It is not a fine-tuning experiment. Comparing the two reports illustrates different learning processes, not a controlled comparison of architecture quality.