| Popis: |
Transfer learning is the advancement over conventional machine learning in which we transfer the newly obtained knowledge to existing knowledge. Traditional machine learning makes a basic assumption that the distribution of training data and testing data should be the same. But in numerous real-world cases, this identical-distribution assumption of training data and testing data does not hold at all. For example, suppose if we have a model to recognize a face from an image in traditional machine learning, we cannot retrain this model to detect tumors in the brain because they belong to a different domain. But using transfer learning, we can retrain this model to detect tumors as well. The identical-distribution assumption might be violated in cases where data from one new domain comes, while there are only available labeled data from a similar other domain. Labeling the new data in the old domain can be costly for any organization, and it is also inappropriate to throw away the newly obtained data just because it is from another domain. |
