03 / 15 PROJECT
Every projectAlexNet Transfer-Learning Image Classifier
A transfer-learning study on AlexNet: freeze some layers, retrain others, and measure what each strategy actually changed.
- AlexNet
- pretrained, partly thawed
- PyTorch
- python, cnns
- Mar 2026
- completed
What the study did
AlexNet, already trained, was pointed at a new classification task. The interesting variable was not the accuracy at the end but which parts of the network were allowed to move. Convolutional layers were frozen and re-trained selectively, layer by layer, and each result was compared against a full fine-tune of the whole network. Written in Python with PyTorch.
Layer by layer
The analysis went below the headline number. Individual convolutional layers were probed to see what they had learned, and the strategies were set against each other: full fine-tuning on one side, selective re-training on the other. The question was not only which scored better but which layers the accuracy actually came from, which is the part a single number hides.
The interesting variable was not the accuracy at the end but which parts of the network were allowed to move.
What is not here
There is no repository and no demo for this one. It was finished in March 2026. What exists is the study itself, the trained models and the analyses that compare them, rather than something you can open in a browser and click around.
What it does
- 01Fine-tuned AlexNet and probed individual convolutional layers
- 02Compared full fine-tuning against selective layer re-training
- 03Layer-wise and performance analysis of the learned features
- 04Applied transfer learning to a new classification task
Built with
- Python
- PyTorch
- CNNs
- Transfer Learning
- AlexNet
- Computer Vision
- Analysis