Transfer learning is one of the most powerful capabilities in the deep learning toolkit because you only need “small data” as opposed to “Big Data”. It’s a technique that allows us to take a deep neural network trained to solve one task (like recognizing objects and logos in an image), and efficiently tweak it to perform another task, like making recommendations based on whether or not fashion items are likely to match with one another… which is exactly what we did in this demo. The beauty of transfer learning is that it allows you to enjoy the accuracy and flexibility that deep learning brings without having to pay its high set-up cost — namely, training effort.

We partnered with DeepLearning.TV to introduce this concept of transfer learning and show how we used it to build a fashion matching demo with our Custom Collections API. Follow along with us in this series of three posts. In Part 1, we take a closer look at how transfer learning works, and why you can’t use it with traditional machine learning algorithms.

 
Check out Part 2 of this series to learn more about about the fashion matching demo we built using Custom Collections! Part 3 is a tutorial walking you through how to build the demo yourself.

About Custom Collections

Our Custom Collections API allows you to build customized machine learning algorithms with about 100x less data compared to starting from scratch. Check out our docs to learn more!

P.S. We’re also holding a competition this month where you can win 100k API calls (and more) for building an image classifier with Custom Collections!

Suggested Posts

Analyzing the Republican Presidential Candidates with Machine Learning

How to Visualize Email Sentiment with Python

Deep Advances in Generative Modeling