Lately, there are lots of posts on one-shot learning. I tried to figure out what it is by reading some articles. To me, it looks like similar to transfer learning, in which we can use pre-trained model weights to create our own model. Fine-tuning also seems a similar concept to me.
Can anyone help me and explain the differences between all three of them?
They are all related terms.
From top to bottom:
One-shot learning aims to achieve results with one or very few examples. Imagine an image classification task. You may show an apple and a knife to a human and no further examples are needed to continue classifying. That would be the ideal outcome, but for algorithms.
In order to achieve one-shot learning (or close) we can rely on knowledge transfer, just like the human in the example would do (we are trained to be amazing at image processing, but here we would also exploit other knowledge like abstract reasoning abilities, and so on).
This brings us to transfer learning. Generally speaking, transfer learning is a machine learning paradigm where we train a model on one problem and then try to apply it to a different one (after some adjustments, as we'll see in a second).
In the example above, classifying apples and knives is not at all trivial. However, if we are given a neural network that already excels at image classification, with super-human results in over 1000 categories... perhaps it is easy to adapt this model to our specific apples vs knives situation.
This "adapting", those "adjustments", are essentially what we call fine-tuning. We could say that fine-tuning is the training required to adapt an already trained model to the new task. This is normally much less intensive than training from scratch, and many of the characteristics of the given model are retained.
Fine-tuning usually covers more steps. A typical pipeline in deep learning for computer vision would be this:
Note the head of our model does not match our needs (there's probably one output per category, and we only need two categories now!)
Swap the very last layer(s) of the model, so that the output matches our needs, but keeping the rest of the architecture and already trained parameters intact.
Train (fine-tune!) our model on images that are specific to our problem (only a few apples and knives in our silly example). We often only allow the last layers to learn at first, so they "catch up" with the rest of the model (in this case we talk about freezing and unfreezing and discriminative learning rates, but that's a bit beyond the question).
Note that some people may sometimes use fine-tuning as a synonym for transfer learning, so be careful about that!
