It is a sub-field of Machine Learning research, focusing on algorithms capable of learning from few data. By using the differences between images rather than the images individually, these algorithms learn to recognize novel objects from less than a dozen examples (sometimes even one!).
Few-Shot Learning has quickly become a key topic in Machine Learning research. Due to their light weight and efficiency, chances are that these methods will become the reference in the AI industry in the next few years. But why wait?
First Siamese Networks
Our first Few-Shot algorithm in production
Dedicated Sicara researchers
An exponential growth
A central challenge is to explain these two aspects of human-level concept learning: How do people learn new concepts from just one or a few examples? And how do people learn such abstract, rich, and flexible representations?
When can a Few-Shot algorithm be the best fitted AI solution for your problem? - You do not have thousands of images at your disposal; - You are able to obtain that many images, but do not want to wait for them before you get first working results; - Your problem involves the recurrent integration of new types of objects. You want to know whether our specific skills on this topic can help you achieve your goals? Reach out to one of our data scientists.
Rather than watching Few-Shot Learning research go by, we chose to be a part of it. We partnered with the engineering school CentraleSupélec. Our very own PhD candidate Etienne is preparing a thesis on interpretable meta-learning algorithms applied to few-shot object recognition. Our goal: help pushing forward the state of the art in Few-Shot Learning. Day by day, everything we learn is helping us to bring value to our clients.
In a production line, we cannot afford to send a piece of machinerie back to inventory every time it needs a fix. Thus a key challenge is for the operators to be able to autonomously maintain heavy machinerie. Once a maintainer has zeroed in on the defective part, they need to identify the part to access its manual, or to order a replacement if need be. With our client, we developed a solution based on a mobile app. Once the maintainer has taken a picture of a part, our algorithm identifies this part among tens of thousands of instances. In one second, it prompts the part's ID and its picture in the catalog. Thanks to few-shot learning, our algorithm only need one example image for each part, and our client can add or remove elements to the catalog without the need for any intervention from a data scientist. We achieved 95% classification accuracy in 4 weeks.
Our articles on Few-Shot Learning
Few-Shot Image Classification with Meta-Learning
Here is how you can teach your model to learn quickly from a few examples.
Your Own Few-Shot Classification Model Ready in 15mn with PyTorch
Always wanted to use state-of-the-art algorithms for your project, but never knew where to start? We got you.