Transfer Learning for Deep Learning with CNN

1. Transfer Learning – Objective

In this blog, we will study Transfer Learning. As this Transfer Learning concept relates with deep learning and CNN also. Although, will use graphs and images to understand Transfer Learning concept.

Introduction to Transfer Learning

Introduction to Transfer Learning

2. Introduction to Transfer Learning

We can say transfer learning is a machine learning method. In this, a model developed for a task that was reused as the starting point for a model on a second task.
Introduction to Transfer Learning

Introduction to Transfer Learning

Transfer learning is the most popular approach in deep learning. In this, we use pre-trained models as the starting point on computer vision. Also, natural language processing tasks given the vast compute and time resource. Although, we need to develop neural network models.
As transfer learning is related to many problems. Such as multi-task learning and concept drift. Although it is not exclusively an area of study for deep learning.
Introduction to Transfer Learning

Introduction to Transfer Learning

Introduction to Transfer Learning

Introduction to Transfer Learning

3. What is a Pre-Trained Model?

To solve a problem, we need to have a pre-trained model of similar problem. Instead of building a model from scratch to solve a similar problem, we use the model trained on other problem as a starting point.

4. How can I use Pre-trained Models?

As there is a predefined aim to use a pre-trained model. Also, a concept of transfer learning plays an important role in a pre-trained model.

While choosing a pre-trained model, one should be careful in their case. If the problem statement we have at hand is very different from the one on which the pre-trained model was trained – the prediction we would get would be very inaccurate.
As already many pre-trained architectures are directly available for use in the Keras library. Imagenet data set has been widely used to build various architectures since it is large enough (1.2M images) to create a generalized model. Although, the problem statement comes in training a model.

That can correctly classify the images into 1,000 separate object categories. Moreover, these 1,000 image categories represent object classes that we come across in our day-to-day lives. Such as species of dogs, cats, various household objects, vehicle types etc.

We use transfer learning to generalize into images outside the ImageNet dataset. This happens only in case of a pre-trained model. Also, we use fine-tuning model for the modifications in a pre-trained model. Since we assume that the pre-trained network has been trained quite well. thus, we don’t want to modify the weights too soon and too much. While modifying we generally use a learning rate smaller than the one used for initially training the model.

5. Ways to Fine tune the model

Ways to Fine tune the model

Ways to Fine tune the model

a. Feature extraction – For a feature extraction mechanism, we use a pre-trained model as in this we can remove the output layer. Further, we have to use the entire network as a fixed feature extractor for the new data set.

b. Use the Architecture of the pre-trained model – According to a dataset, at the time of initializing and training model, we use its architecture. 

c. Train some layers while freeze others – There is one more way to use a pre-trained model i.e to train model partially. Further, we have to keep the weights of initial layers of the model frozen. While have to retrain only higher layers.  We can try and test as to how many layers to be frozen and how many to be trained.
The below diagram should help you decide on how to proceed with using the pre-trained model in your case –

Scenario 1 – Size of the Dataset is small while the Data similarity is very high – As in this particular case, we do not require to retain the model, as data similarity is very high. Although, according to our problem statement, we need to customize and modify the output layers. As we have use pre-trained model here as a feature extractor.

Further, to identify the new set of images have cat or dogs, we use trained models on Imagenet. Here we require similar images to Imagenet to categorize the two outputs – cats or dogs. Finally, at last in this case, we have to modify dense layers. Also, have to put the final softmax layers to output 2  categories instead of 1000.

Scenario 2 – Size of the data is small as well as data similarity is very low – As in this case, we have to freeze the initial (let’s say k) layers of the pre-trained model. Also, as freezing complete, then train the remaining(n-k) layers again. Although, keep in mind that the top layers would be customized to the new data set.  Also, initial layers are kept pre-trained by their smaller size. But, keep frozen weights of those layers.

Scenario 3 – Size of the data set is large however the Data similarity is very low – Particularly, in this case, neural network training would be more effective. As it’s having a large data set. Also, the main thing is that the data we use is different. As we use data is different from data we use in training. Hence, its best to train the neural network from scratch according to your data.

Scenario 4 – Size of the data is large as well as there is high data similarity – We can say this is the final and the ideal situation. As pre-trained models are more effective in this case. Also, we can use this model in very good manner. We have to just use the model is to retain the architecture of the model and the initial weights of the model. Moreover,  we can retrain this model using the weights as initialized in the pre-trained model.

6. Inductive learning and Inductive Transfer

We use this form of transfer learning in the deep learning, called for an inductive transfer. Thus, it’s an area where the scope of possible models is narrowed in a beneficial way. Although, this model fit into a different but related task.
Inductive learning and Inductive Transfer

Inductive learning and Inductive Transfer

7. How to Use Transfer Learning?

Two common approaches for transfer learning are as follows:
  • Develop Model Approach
  • Pre-trained Model Approach

a. Develop Model Approach

Select Source Task:
While selecting a task, we must have to select predictive modeling problem. The problem with an abundance of data.
Develop Source Model:
Next, we have to develop a skilful model for this first task. Although, a model must be better than the naive model. That is to ensure some of the model must be better than a naive model.
Reuse Model:
The model must fit on the source task. That further we use it as the starting point for a model on the second task of interest. Moreover, this involves parts of the model, depending on the modeling technique used.
Tune Model:
We need to adopt model on the input-output pair data available for the task of interest.

b. Pre-trained Model Approach

Select Source Model:
We have to choose a pre-trained source model from available models. Models are released on large and challenging datasets by many research institutions.
Reuse Model:
As on the starting point, we can use pre-trained model. That is the second task of interest. Depending on the model used, it involves all parts of the model.
Tune Mode:
We need to adopt model on the input-output pair data available for the task of interest.

8. When to Use Transfer Learning?

We use transfer learning to save time or for getting better performance as it is an optimization.
There are three possible benefits to look for when using transfer learning:
Higher start:
The initial skill on the source model is higher than it otherwise would be.
 
Higher slope:
The rate of improvement of skill during training of the source model. That is steeper than it otherwise would be.
Higher asymptote:
The converged skill of the trained model is better than it otherwise would be.
Transfer Learning for Deep Learning with CNN

Transfer Learning for Deep Learning with CNN

9. Conclusion

As a result, we have studied Transfer Learning. Also, learned all W’s of Transfer Learning. Along with this, we have studied concepts with diagrams. Furthermore, if you feel any query, feel free to ask in a comment section.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.