Skip to main content
Version: v0.3.0

Zero-shot learning models

Overview​

By default, H2O Label Genie utilizes certain zero-shot learning models to accelerate the labeling process. In particular, H2O Label Genie lets you use a zero-shot learning model for the following supported annotation tasks:

A zero-shot learning model can be turned On or Off at any time. To learn more, see Smart annotation.

What are zero-shot learning models?​

Labeled data is crucial for supervised learning problem types in computer vision (CV), natural language processing (NLP), and audio. High-quality labeled data usually requires a lot of manual labeling that can lead to high costs and delay production or execution.

One way to accelerate the labeling process is to utilize zero-shot learning models. These models let data scientists label unlabeled data with high accuracy and speed. Zero-shot learning models are pre-trained models that have been trained on vast and distinct classes. As a result, zero-shot learning models with prior knowledge can label unlabeled data.

What are zero-shot learning model predictions?​

The labels or suggested labels for a given sample that are provided by a zero-shot learning model are called zero-shot learning model predictions. For example, for an image and text classification annotation task, H2O Label Genie, with a zero-shot learning model activated, offers a percentage probability of an image or text belonging to a certain label (class). For an object detection annotation task, it populates an image with bounding boxes where the bounding boxes capture the desired objects (for example, a car).

Annotation tasks + zero-shot learning models​

Image classification​

For image classification annotation tasks, H2O Label Genie utilizes the OpenCLIP zero-shot learning model. OpenCLIP is an adaptation of OpenAI's Contrastive Language-Image Pre-training (CLIP).

Object detection​

For object detection annotation tasks, H2O Label Genie utilizes the Detic zero-shot learning model.

Image instance segmentation​

For image instance segmentation annotation tasks, H2O Label Genie utilizes the Detic zero-shot learning model.

Text classification​

For text classification annotation tasks, H2O Label Genie utilizes the bart-large-mnli zero-shot learning model.

Text summarization​

For text summarization annotation tasks, H2O Label Genie utilizes the following zero-shot learning models: bart-large-cnn, distilbart-cnn-12-6, and pegasus-large.

Feedback