Zero-Shot Learning vs. Few-Shot Learning vs. Fine Tuning
A quick understanding of the key differences between various techniques for adapting Large Language Models
Note — these techniques are not primarily for LLMs, they can also be used for image classification, speech recognition, recommendation systems and other tasks.
Introduction
Large Language Models (LLMs) like GPT-3 and BERT have transformed how computers understand and generate human-like text. These models can perform many tasks, from writing essays to answering questions, without needing specific instructions for each task. The way these models learn and adapt to new tasks can be divided into three main techniques: zero-shot learning, few-shot learning, and fine-tuning.
Zero-Shot Learning
A technique where it allows a model to perform without any labeled examples, only classifying based on what it has learned from general data.
Zero-shot learning allows a model to perform a task without being specifically trained for it. For example, a model can summarise a paragraph, answer trivia questions, or translate text just based on its general understanding of language from reading a lot of text during its training. It doesn’t need any examples to perform these tasks.
Imagine a model that has never been explicitly trained on sentiment analysis. However, it has been trained on a vast amount of text covering various topics and emotions. With zero-shot learning, you can ask the model to determine the sentiment of a new piece of text, such as whether a given review is positive or negative, without any additional examples.
The model uses its general knowledge of language and sentiment to make a prediction. Just as a well-read person might use their understanding of emotional expressions to interpret a new piece of text, the model uses its broad training to make an educated guess about the sentiment.
Benefits —
Flexible and cost-effective because it doesn’t need any extra training data.
Downsides —
Usually less accurate than models trained for specific tasks and might not perform well on specialised or complex topics. This is due to the model’s performance depending heavily on how much and what type of data it was initially trained on.
Few-Shot Learning
A technique where it allows a model to perform with several concrete examples of task performance.
Few-shot learning is when a model is given a few examples to learn a new task. Unlike zero-shot learning, it uses these examples to better understand what’s being asked. For instance, you can give a model a few examples of questions and answers, and it can use this information to answer similar new questions more accurately.
Now, let’s say you want the model to classify text as either “Spam” or “Not Spam” . The model is provided with a few example sentences labeled as spam or not spam, such as “Congratulations! You’ve won a free prize. Click here to claim your reward!” labeled as “Spam” and “Meeting agenda for tomorrow’s team discussion” labeled as “Not Spam.” With these few examples, the model learns the features and patterns that distinguish junk messages from regular ones. So, when it encounters a new message, it uses these examples to make a more accurate classification.
The model achieves this as few-shot learning models are often designed with mechanisms such as meta-learning. Meta-learning trains models in a way that they not only learn from the specific examples provided but also learn how to adapt their learning process for new tasks. This involves learning general strategies or patterns from various tasks and then quickly adapting to a new task with just a few examples. Instead of relying on large datasets, this “learning to learn” method involves creating models that are flexible and can generalise knowledge from previous tasks or examples to new ones.
Benefits —
Often performs better than zero-shot learning since it has some context to work with and it still doesn’t require a lot of data.
Downsides —
Depends on the quality of the examples provided; if the examples are poor, the model’s performance will suffer.
Fine-Tuning
A technique that allows a model to be retrained on a variety of concrete examples, and updates with new knowledge, resulting in a newly refined version.
Fine-tuning involves taking a general-purpose model and training it further on a specific set of data related to a particular task. This makes the model highly specialized and accurate for that task.
Imagine you have a large language model that has been trained on a broad range of general text data. Now, you want this model to excel at analysing legal documents, such as contracts or court rulings. To achieve this, you perform fine-tuning. You start with the pre-trained model and further train it on a specialised dataset consisting of thousands of labeled legal documents. This dataset includes various types of legal texts, such as contract clauses labeled with their type (e.g. “Confidentiality Clause,” “Termination Clause”) or court rulings labeled with the legal outcomes (e.g. “Guilty,” “Not Guilty”).
After this fine-tuning process, the model becomes highly specialised in understanding and classifying legal language. When you input a new legal document, the fine-tuned model can accurately identify and categorise different clauses or legal outcomes based on its specialised training. Fine-tuning adjusts this model’s parameters to better handle the specific nuances and terminology of legal documents.
Benefits —
High accuracy and reliability for specific tasks, making it ideal for important applications.
Downsides —
It requires a large amount of data, can be expensive and time-consuming to train. The model becomes less flexible and might not work well on tasks outside its specialised training.
