This is my 2nd post in this series of ‘Navigating the vocabulary of Gen AI’, and in this post I continue and follow on from the first post I made here where I provided an overview of the following AI terminology:
- Artificial Intelligence
- Machine Learning
- Artificial Neural Networks (ANN)
- Deep Learning
- Generative AI (GAI)
- Foundation Models
- Large Language Models
- Natural Language Processing (NLP)
- Transformer Model
- Generative Pretrained Transformer (GPT)
Responsible AI is designed to set out the principles and practices when working with artificial intelligence to ensure that it is adopted, implemented and executed fairly, lawfully, ethically ensuring trust and transparency is given to the business and its customers. Considerations to how AI is used and how it may affect humanity must be governed and controlled by rules and frameworks. Trust, assurance, faith and confidence should be embedded with any models and applications that are built upon AI.
Labelled data is used to help machine learning models and algorithms process and learn from raw material. The data is ‘labelled’ as it contains tags and features associated with the target data which provides useful and informative information about it, for example if you had a photo of a tiger, it could be labelled with ‘Tiger’. This helps to provide context to the raw data which the ML model can then use and extract to help it to learn and recognise other images of tigers. This raw input data can be in the form of text, images, videos and more and requires human intervention to label the data correctly.
Supervised learning is a training method used within machine learning which uses a vast amount of labelled datasets in order to be able to predict output variables. Over time, the algorithms learn how to define the relationship between the labelled input data and the predicted output data using mapping functions. As it learns, the algorithm is corrected if it makes an incorrect output mapping from the input data, and therefore the learning process is considered to be ‘supervised’. For example, if it saw a photo of a lion and classified it as a tiger, the algorithm would be corrected and the data sent back to retrain.
Unsupervised learning differs from supervised learning in that supervised learning uses labelled data, and unsupervised learning does not. Instead it is given full autonomy in identifying characteristics about the unlabeled data and differences, structure and relationships between each data point. For example, if the unlabeled data contained images of tigers, elephants and giraffes, the machine learning model would need to establish and classify specific features and attributes from each picture to determine the difference between the images, such as colour, patterns, facial features, size and shape.
This is a method of learning that uses a combination of both supervised and unsupervised learning techniques and so uses both labelled and unlabeled data in its process. Typically when using this method, you have a smaller data set of labelled data compared to a larger data set of unlabelled data, this prevents you having to tag a huge amount of data. As a result this enables you to use the smaller set of supervised learning to assist in the training of the model and so aids in the classification of data points using the unsupervised learning technique.
Prompt engineering allows you to facilitate the refinement of input prompts when working with large language models to generate the most appropriate outputs. The technique of prompt engineering enables you to enhance the performance of your generative AI models to carry out specific tasks by optimising prompts. By making adjustments and alterations to input prompts you can manipulate the output and behaviour of the AI responses making them more relevant. Prompt engineering is a principle that is allowing us to transform how humans are interacting with AI.
Prompt chaining is a technique used when working with large language models and NLP, which allows for conversational interactions to occur based on previous responses and inputs. This creates a contextual awareness through a succession of continuous prompts creating a human-like exchange of language and interaction. As a result, this is often successfully implemented with chat-bots. This enhances the user’s experience by responding to bite-sized blocks of data (multiple prompts) instead of working with a single and comprehensive prompt which could be difficult to respond to.
Retrieval augmented generation (RAG)
RAG is a framework used within AI that enables you to supply additional factual data to a foundation model as an external source to help it generate responses using up-to-date information. A foundation model is only as good as the data that it has been trained on, and so if there are irregularities in your responses, you can supplement the model with additional external data which allows the model to have the most recent, reliable and accurate data to work with. For example, if you asked ‘what’s the latest stock information for Amazon’ RAG would take that question and discover this information using external sources, before generating the response. This up-to-date information would not be stored within the associated foundation model being used
AI parameters are the variables within a machine learning model that the algorithm adjusts during training to enable it to optimise its performance to generalise the patterns from data, and therefore making them more efficient. These values dictate the model’s behaviour and minimise the difference between predicted and actual outcomes.
Fine-tuning is the technique of adjusting a pre-trained model on a particular task or data set to improve and enhance its performance. Initially trained on a broad data set, the model can be fine-tuned using a smaller, and more task-specific data set. This technique allows the model to alter and adapt its parameters to better suit the nuances of the new data, improving its accuracy and effectiveness for the targeted application.
In my next post I continue to focus on AI, and I will be talking about the following topics:
- Emergence in AI
- Text Classification
- Context Window