Resume (What did I do in a nutshell ?)
We first configured a basic model with a succession of specific layers, in order to be able to learn from a database of images coming from a public database (ciphar). We have created a model derived from this basic model. We trained this model for the first time in freeze mode. Then a second time, in unfreezed mode. This step has inherited the learning from the first pass.
Introduction (What is the problem ?)
Humans do not learn without knowledge. They base themselves on their knowledge already acquired in order to optimize their learning. They learn for example to recognize an animal, then as soon as it is acquired they manage to recognize another animal based on their previous acquired ability. Babies are special, because we cannot say that they already have a knowledge base to begin their learning. This base comes from their parent. Why am I telling you about this? For machine learning, it would be nice to optimize learning. If one learns to recognize an image of a dog, one should be able to reuse this knowledge to determine the image of a cat.
Materials and methods
First, I documented myself (using the references provided, then I went on the internet and in books to understand what transfer learning means). Every time I found something, I belived to understand something, I tried to put it into practice.
Results
Blue curves represent training data.
Orange curves represent validation data.
Discussion (What does it mean ?)
We can see that at each epoch, the precision increases while the loss decreases. In the last epoch, that of the final model, we can see that the precision is higher than the precision of the last epoch of the base model (the loss is also lower than that of the last epoch of the base model). It can therefore be seen that the final model inherits the prior learning of the base model.
Acknowledgements (Who helped me out ?)
Thanks to Nathan LAPEYRE, with whom I was able to work and with whom I was able to confront my ideas step by step for this study.
