Skip to main content

Transfer Learning

Build upon pre-existing knowledge, streamline the learning process for novel tasks

Input: Images, videos, and texts

Output: Concepts

A transfer-learning model, which we previously called an "embedding-classifier", is a type of machine learning model that uses an embedding layer to represent images, videos, or texts as low-dimensional vectors, and then uses a classifier layer to predict the class of the input.

The embedding layer takes the input image, video, or text, and converts it into a vector of numbers, where each number represents the similarity of the input to a particular word or concept. The classifier layer then takes the embedding vector as input and predicts the class of the input.

A key advantage of transfer-learning models is that they can be trained on large datasets of images, videos, or texts without requiring a lot of computational resources. This is because the embedding layer can be pre-trained on a large dataset of images, videos, or texts, and then the classifier layer can be trained on a smaller dataset of labeled images, videos, or texts.

Note: Essentially, transfer learning leverages the knowledge gained from a pre-trained model to facilitate the learning process of a new model for a related problem. The pre-trained embedding model serves as a feature extractor that has already learned useful features from a large dataset. This knowledge is transferred to the classifier layers, enabling effective classification with minimal training data.

You may choose the transfer learning model type if you want to:

  • Seamlessly transfer the knowledge gained from existing models to solve problems without necessarily having to train a new model from scratch.

  • Take advantage of a model that has been trained on a similar task, such as object recognition in images. Then, fine-tune this pre-trained model to recognize the specific classes that align with your objectives.

  • Get results in seconds, not minutes or hours — allowing you to achieve significant progress with minimal training data. You do not need many inputs to get started creating a custom model using our transfer learning technology. You can even start with 10 inputs and add more as needed.

© 2025 Clarifai, Inc. All rights reserved