From the course: Introduction to Transformer Models for NLP
Join today to access over 24,800 courses taught by industry experts.
Few-shot learning
From the course: Introduction to Transformer Models for NLP
Few-shot learning
- Section 7.4, "Few-Shot Learning." So the idea of few-shot learning, or one-shot learning, or even zero-shot learning, stems from the fact that GPT, being pre-trained on that massive web text corpora, was actually seeing natural instances of people performing what we would consider as natural language processing tasks. So when we talk about few-shot learning, or zero-shot, or one-shot, is we're asking the model to perform a task that it may have never actually seen before, given some prompt or context into what that task is. Few-shot learning means that we provide the model with a description of the task and as many examples as we desire or would fit in the actual window for the model. For GPT-2, the max input length is 1,024 tokens. Zero or one-shot learning is a subset of few-shot where the number of examples that we give are either zero or one. Zero-shot learning is when we would give a model a task description, but no examples prior. And the prompt is what's used to tell the…
Download courses and learn on the go
Watch courses on your mobile device without an internet connection. Download courses using your iOS or Android LinkedIn Learning app.