Google uses a training image augmentation algorithm to improve model performance in cases of insufficient data and/or incomplete or imperfect labels. The Augmented Learning for Image Classification experiment aims to product higher quality models with as-is training data, enabling faster development and reduced labeling costs.
To use Augmented Learning for Image Classification, you provide labeled training data just as you would for any machine learning model, but with an insufficient number of images and/or incomplete or partially inaccurate labels. Augmented Learning for Image Classification gives you API access to two convolutional neural network (CNN) models: one trained with the data as is, and a second trained with augmented learning.
|Apply for access||Private documentation|
This experiment is designed to help customers create robust vision (image) models. It will be most helpful for customers whose training data is limited:
- An insufficient amount of labeled images (<50 per class)
- Image labels that may be noisy (>20% error rate)
It should be effective with image-based challenges in a wide range of industries and functions. Past opportunities have included retail, entertainment, real estate, industrial, and insurance.
What data do I need?
Augmented Learning for Image Classification is likely to be effective with a wide range of image types, including real-world and abstract objects, characters, and so on. It may not be effective with highly unusual image types, such as specialized medical images and scans.
The experiment supports a wide range of label types, including specific elements within the image (such as a dog or flower) as well as characteristics of the image (such as color, shape, texture, sentiment, quality).
- Users must provide at least five images across two or more classes of imperfectly-labeled data
- Images may be stored in one of the following formats: JPEG, PNG, WEBP
- Images are labeled in a CSV format
What skills do I need?
As with any AI Workshop experiment, successful users are likely to be savvy with core AI concepts and skills in order to both deploy the experiment and interact with our AI researchers and engineers. In particular, users of this experiment should:
- Be familiar with the Google Cloud ML Engine API for model inference
- Understand the basics of classification problems