Transfer Learning with TensorFlow

Transfer Learning with TensorFlow session in HacktoberFest Ahmedabad 2020

Transcript

2. What is Transfer Learning It’s a technique in ML field where a model trained on one task is reused for similar or related task It improves performance of second task modeling It’s not just for images but can be used for many other areas
3. Traditional Machine Learning Task 1 / Domain 1 Model 1 Result Task 2 / Domain 2 Model 2 Result Training of same domain or related task
4. Transfer Learning Task 1 / Domain 1 Model 1 Result Knowledge Sharing knowledge gained from one task to related task Model 2
5. Machine Learning’s Next Frontier • Trasnfer learning will be the next driver of ML Success • So far many models are trained with high accuracy and good performance • Trained data is now available for many tasks and domains • Now it’s time to apply knowledge from this models to related tasks and domain • Expectations from end user is that model should perform many related tasks rather than the one on which it’s trained • Learn from many experiences and export knowledge to new environments
6. Transfer Learning Approach • Develop Model Approach • Pre Trained Model Approach
7. Develop Model Approach • Source Task – Select a problem where plenty of data is available and have a relationship between input and output data. • Develop a Model - Develop a model for above task and ensure that some feature learning is performed • Reuse Model – This model can be reused as starting point for related tasks • Tune Model – Model can be tuned for related tasks to perform better
8. Pre Trained Model Approach • Source Source Model – Pre trained model is selected from related models • Reuse Model – This model can be reused as starting point for related tasks • Tune Model – Model can be tuned for related tasks to perform better
9. Applications of Transfer Learning • Learning from simulations • Transfer Learning in Computer Vision and CNN • Transfer Learning in NLP • Multi Tasking
10. Transfer Learning Strategies • What to Transfer – Identify the features that can be transfered • When to Transfer – Making sure that transfer does not degrade the peformance • How to Trasnfer – Identify the ways to transfer knowledge to related tasks
11. Transfer Learning Strategies • Inductive Transfer Learning • Same Source and Target Domain • Tasks are different • Labeled data is avaialble • Unsupervised Transfer Learning • Same Source and Target Domain • Tasks are different • No labeled data • Transductive Transfer Learning • Different Source and Target Domain • Tasks are similar • Source domain has labeled data
12. TensorFlow • It’s an end to end open source machine learning platform • Easy Model Building • Developed to run large numerical computations • High level APIs • Supports Deep Learning, Nueral Networks etc.
13. Coding Time Transfer Learning in NLP with TensorFlowHub
14. Datasets Available for Transfer Learning • Google’s word2vec Model • Stanford’s GloVe Model • FastText • Oxford VGG Model • Google Inception Model • Microsoft ResNet Model • ImageNet Model
15. Thank You Q & A Twitter : https://twitter.com/hjdave LinkedIn: https://linkedin.com/in/davehiren Blog : https://hirendave.tech https://hiren.dev