Transfer Learning is the technology behind Microsoft’s Custom Vision service, which enables the creation of custom vision models with an extremely modest amount of training data.

However, this is just the tip of the technology’s iceberg.

Transfer learning is the reuse of a pre-trained model on a new problem. It’s currently very popular in deep learning because it can train deep neural networks with comparatively little data. This is very useful since most real-world problems typically do not have millions of labeled data points to […]

Transfer learning is a statistical technique that’s been getting more attention lately that enables you to reuse a model for a different task than what it was trained for. In this episode, Siraj Raval shows you how to use transfer learning to predict instances of gold deposits using publicly available satellite imagery.

I’m glad to see that the Azure Custom Vision Service is getting some press. It’s an easy and simple way to build your own computer vision models without having to train on thousands (or tens of thousands) of images. In fact, as little as 15 images can yield workable results.

Here’s an article in www.itbusiness.ca about the service.

“Customers can train their own custom image classifiers and object detectors,” said Tina Coll, the product marketing manager at Microsoft Corp. “For example, a company could choose to detect their own logo in the video of a sports event to track the impact of their advertising or a student might want to count the number of animals passing in front of a nature camera.”

Angel Wong demonstrates the AP2 (AutoPilot V2) features of her new Tesla 3. It’s not quite autonomous driving, but it is really close. I wonder how long it will be before consumers will come to expect this feature on all new cars.

Of particular interest to me is the “training period” where Autopilot is not available as the car is learning about your normal environment and your driving habits. Sounds a bit like a little transfer learning is going on.

By leveraging powerful prior knowledge about how the world works, humans can quickly figure out efficient strategies in new and unseen environments.

Currently, even state-of-the-art Reinforcement Learning algorithms typically don’t have strong priors and this is one of the fundamental challenges in current research on Transfer Learning.

Related Links