Finding a simpler path to better computer vision

Computer vision simplified: A simple path to improved computer vision

A machine-learning algorithm must first be trained before it can perform a specific task, like identifying cancerous cells in medical images. Image classification models are typically trained by showing them millions of images in a large dataset.

Real image data, however, can raise ethical and practical concerns. The images may violate copyright laws, violate privacy rights, or even be biased towards a particular racial group or ethnicity. Researchers can create synthetic data using image generation software to avoid these pitfalls. These techniques are not as effective because they require expert knowledge to design an image generation program which can produce useful training data.

Researchers at MIT and the MIT IBM Watson AI Lab took a new approach. They gathered 21,000 freely available internet programs instead of creating customized image-generation programs for a specific training task. They then used this collection of basic image-generation programs to train a model for computer vision.