Hi! I didn’t understand this topic, honestly. So, we trained model, where 10 examples of signature are classified as similar. Customer of bank added 1 signature, that differenciate from this 10 examples. So, you say, that we can say to model, that this 11 signatures are similar without retraning. How is it?
This kind of learning means that given a pretrained model for a particular task, you continue to train it further (maybe in a cycle that runs several times) with a few other examples (inputs and labels given).
Then at the end of this extended training process the model should be able to correctly classify examples similar to the ones used in the extended training.
hmm, so, in this case we should retrain model, isn’t it?
Yes it get retrained or better say fine tuned.
so, we can retrain any models, isn’t it? What is the uniquness of One Shot Learning?
You dont retrain for a long time thats its uniqueness, otherwise you would time and more resources to retrain.
and how is it achieved?
The context of the one shot learning dirrcts the LLM to provide a better answer.
do you mean, that we shouldn’t train from zero with passing of all data? And this weights trained earlier you named “context”, right?
To train from scrarch is a very “big job” you need a lot of data and computing resources.
With a few shot techniques, the weights do not change, just the LLM is directed towards a better context.
what is the difference with transfer learning in this case?
I understood the concept differently. In my understanding the model during training learns to distinguish one signature from another. So if a new customer is added , there will be a need to compare a check he/she has signed with the signature he has provided at the time of opening the account. The model is not required to be able to distinguish the customer from signature but that the signature provided by the customer on bank opening matches the check he/she has provided are same or not. There should be no need to re-train the model for that. The model is already trained to calculate the difference.
I am not sure if I am correct, but that is what I could make of it