Chapter 3: Harnessing the Power of Pre-Trained Networks with Transfer Learning
Despite the undeniable power deep neural networks bring to computer vision, they are very complex to tune, train, and make performant. This difficulty comes from three main sources:
- Deep neural networks start to pay off when we have sufficient data, but more often than not, this is not the case. Furthermore, data is expensive and, sometimes, impossible to expand.
- Deep neural networks contain a wide range of parameters that need tuning and can affect the overall performance of the model.
- Deep learning is very resource-intensive in terms of time, hardware, and effort.
Do not be dismayed! With transfer learning, we can save ourselves loads of time and effort by leveraging the rich amount of knowledge present in seminal architectures that have been pre-trained on gargantuan datasets, such as ImageNet. And the best part? Besides being such a powerful and useful tool, transfer learning...
