Generative AI For The Rest Of US

Few-Shot Learning

Zero-Shot Learning and Few-Shot Learning (FSL) are both designed to enhance the applicability of AI models to tasks or data they haven't explicitly been trained on. However, they serve different purposes and are suited to different scenarios. Since ZSL relies on the model's existing knowledge and generalization capabilities, it's more suitable for tasks where the model can infer or generalize easily from its training data. However, when the task requires a closer alignment with the new use cases, FSL can be more effective.

Also, in some scenarios where collecting large datasets is impossible, FSL could provide a way to achieve significant improvements in model performance without the need for extensive data collection. Indeed, FSL is essentially a balance between the broad generalization capabilities and the need for a more specific set of examples to guide the model's responses.

Generally, FSL is applicable across various fields. In medical imaging, like MRI or X-ray analysis, it addresses the challenge of limited data for new or rare diseases. A model trained on broad medical imaging datasets can be adapted to detect rare disease signs with a few annotated images. For example, with a minimal set of X-rays of a rare lung condition, FSL can train the model to identify this condition in new patients.

In practice, when using LLMs like ChatGPT, we can leverage FSL by providing a few examples of the input-output pairs we want the model to understand. To demonstrate a simple scenario, we can use ChatGPT and input a few examples then observe how the model performs on new instances.

This is the prompt to copy and paste into ChatGPT:

User: Paris
Assistant: France, earth, solar system, the Milky Way, Local Group, Virgo supercluster, observable universe

User: London
Assistant: United Kingdom, earth, solar system, the Milky Way, Local Group, Virgo supercluster, observable universe

User: Tokyo
Assistant: Japan, earth, solar system, the Milky Way, Local Group, Virgo supercluster, observable universe

User: Berlin

Submit and observe how the model performs on the prediction of the next location in the sequence as a response to the last user prompt (Berlin). The expected response is:

Assistant: Germany, earth, solar system, the Milky Way, Local Group, Virgo supercluster, observable universe