Can we take our model further? Can we achieve close to 90%, reaching human level performance? As we shall see in this section, we can obtain better performance by leveraging on transfer learning.
Transfer learning is a technique in machine learning where a model trained for a certain task is modified to make predictions for another task. For example, we may use a model trained to classify cars to classify trucks instead, since they are similar. In the context of CNN, transfer learning involves freezing the convolution-pooling layers, and only retraining the final fully connected layers. The following diagram illustrates this process:
How does transfer learning work? Intuitively, the ...
