Introducing Transfer Learning as Your Next Engine to Drive Future Innovations
by Harley Davidson Regua
As more and more enterprises realize the advantages deep learning brings, AI enthusiasts begin to buckle up for take-off towards the search for state-of-the-art technologies as the holy grail of the future. Deep learning has become a powerful tool for potential innovations thanks to the massive volume of accessible resources that make AI development simple.
As a curious leader, you might have started your journey into recreating the future of AI and caught yourself jumping on the deep learning bandwagon. With genius data scientists and machine learning developers on your side, you imagine yourself up on the pedestal ready to change the tides of how technology impacts the lives of millions.
Now, you have thought of a use-case for a healthcare industry specific to computer-aided diagnosis that no one has ever done before. You have prepared necessary resources like computing devices and collected a relatively good dataset, however, you found yourself lost in the wilderness of neural network’s hyperparameters with no idea where to start and what to configure.
Good thing is, you don’t really need to start building your neural network from scratch. You now can leverage from the power of past powerful deep learning models and transfer their knowledge to your own specific domain — thanks to transfer learning.
According to Andrew Ng in his tutorial about “Nuts and Bolts of Building AI Applications using Deep Learning” during the 2016 Conference on Neural Information Processing Systems (NIPS 2016), transfer learning will be the next driver of machine learning’s commercial success superseding its predecessor, supervised learning.
The key motivation for the sprout of transfer learning was to overcome the isolated learning paradigm since conventional learning and deep learning algorithms were only designed to solve separate specific tasks, which inconveniently requires a huge amount of data. Besides, acquiring labeled data to be trained for supervised learning adds yet a further layer of complexity. Transfer learning takes advantage of the power of pre-trained models to solve new problems even with relatively few dataset.
So now you have classified your healthcare industry use-case into a predictive modelling problem with sets of medical images as your data input. For this problem, one common approach is to use a deep learning model pre-trained for a large and challenging image classification task such as the ImageNet 1000-class photograph classification competition and fine tune it with your own hyperparameter configuration.
Recent researches in medical imaging have demonstrated the potential of pre-trained convolutional neural networks (CNN) as medical image feature extractors. Based on the study of Lopez and Valiati in 2017,using pre-trained CNNs as feature extractors for tuberculosis detection does not require an expensive and time-consuming training step nor a large dataset to achieve reasonably good results.
Transfer learning has shown its value as a quick solution to pertinent problems related to deep learning. Yet, it requires more research and exploration to address bigger issues surrounding existing learning algorithms like the difficulty of answering what, when, and how to transfer, as well as major challenges of negative transfer and transfer bounds.
Get your FREE Robotic Process Automation (RPA) Software now! Download here
References:
