A major challenge in transfer learning is determining whether the initial and current problems are similar enough. There is no strict determinant for what makes a source network sufficient or similar enough to the target application. If the algorithm fails to find a connection between the problems, the model can perform even worse than if it had never been trained in the first place.

Task Similarity Challenge in Transfer Learning

- Need very large dataset
- Data must be general and not domain-specific
- Only works if initial and current problem are similar enough


University of Michigan - Ann Arbor

The longest part of the development of a deep learning model is that of the training stage. However, if the problem that is being tackled has already been studied and worked on by someone else and the model has been released to the public including the trained weights then a technique to substantially increase the training process is to utilize transfer learning.

Transfer Learning in Deep Learning

https://towardsdatascience.com/transfer-learning-using-the-fastai-library-d686b238213e

Transfer Learning Reference

According to DeepMind CEO Demis Hassabis, transfer learning is also one of the most promising techniques that could lead to artificial general intelligence (AGI) someday.

ideas about transfer learning from the specialist

- Developing your own model
- Using a pre-trained model

Two Approaches to Transfer Learning in Deep Learning

The primary advantages of transfer learning in deep learning include reduced training time, improved neural network performance, and lower data requirements. Since deep neural networks typically require massive datasets and extensive computational time to train from scratch, leveraging a pre-trained model allows for the development of robust models even when domain-specific data is scarce or expensive to acquire.

Advantages of Transfer Learning in Deep Learning

A practical example of fine-tuning involves adapting an image classification model. First, a source model is pre-trained on a large source dataset, such as ImageNet. To adapt this model for a specific task, like classifying images in a smaller target dataset of chairs, a target model is constructed by copying the source model's parameters and replacing its output layer. The new output layer is then trained from scratch on the chair dataset, while the remaining layers are fine-tuned, allowing the target model to leverage the broad visual features learned from the source dataset.

Example of Fine-Tuning for Image Classification

To classify a subset of a large dataset—such as a specific group of classes from ImageNet—it is computationally advantageous to use models pre-trained on the full dataset to extract features. Rather than training a model from scratch or fine-tuning all layers, the pre-trained architecture can be utilized, and only a custom, small-scale output network needs to be trained on the subset. This approach leads to less computational time and memory cost while capitalizing on the robust visual representations already learned by the pre-trained model.

Efficient Subset Classification via Pretrained Feature Extraction

Transfer learning from a source task (Task A) to a target task (Task B) in deep learning is most effective under the following conditions:
1. Tasks A and B share the same input $$x$$.
2. There is significantly more training data available for Task A than for Task B.
3. Low-level features learned from Task A are expected to be helpful for learning Task B.

Learn Before

Related

Learn After

Learn After

Related