Active Learning is an iterative algorithm that selects the most valuable and informative images for labeling. These images can increase the model’s performance, so Data Scientists label them first during the data annotation process. Such an approach saves a lot of resources and leads to faster model development.
Once you have done so, please identify the type of vision AI problem you are solving:
To use the Active Learning feature, you must train the corresponding assistant by labeling 10 images.
If you do not unlock the assistant, a nasty notification will show up:
Different assistants require different types of annotations and tools to produce the needed labels. In our case, our vision AI problem was Object Detection, so we used the Bounding box tool.
Do not forget to set the labeled image’s status as “Done” or “To review.”
When the assistant is unlocked, go to the project dashboard in the burger menu on the left and access the Active Learning feature in the Tools section.
Once you click on it, please enable the Active Learning feature.
You can choose the following heuristics to rank your images:
There are two ways to set the heuristics:
Set the active heuristic for each family model – the Active Learning ranks will be available each time the model retrains;
Generate a rank on demand – the algorithm generates ranks on demand based on the chosen heuristic. In this case, you do not have to wait until the model retrains to get a rank.
In our case, we decided to rank 100 images on demand. It means an algorithm will select 100 images the model is the most uncertain about. The selected images will be the perfect candidates to be labeled next. Once the run is over, we can proceed with the annotation.
Annotate the proposed images, switch their status to “Done” or “To review,” and enjoy the improvement of your assistant.
If you want to learn more about the general Active Learning concept, please check out our article.
Only 13% of vision AI projects make it to production, with Hasty we boost that number to 100%.