Fine-tuning CNNs for Classification

ND320 C2 L3 15 Fine-Tuning CNNs For 2D Medical Image Classification Walkthrough Pt 1

INSTRUCTOR NOTE:

Erratum: When the instructor says "This layer is made up of 128 filters, each with a size of 112x112" , she actually meant the size of output or activations of the filter layer is 112x112. The diagram also shows the shape of the activations while filters are not shown. The filters themselves are 3x3. A full description of a VGG network can be found here

ND320 C2 L3 16 Fine-Tuning CNNs For 2D Medical Image Classification Walkthrough Pt 2

Summary

Fine-tuning

The first several layers of filters trained are only going to learn line- and shape-based features because their visual fields are so small. We can reuse or freeze the pre-trained weights of the first few layers and only need to train filter weights to detect higher-order features that are more relevant to your specific use cases. We call this process that only makes adjustment of weights in the last a few layers fine-tuning.

One of the key pieces of fine-tuning is the last layer. We need to adjust the dimension of the last layer to match our specific use cases. We can also add new layers to train from scratch.