An AI Overview

Before we launch into this...

Allow me to introduce Oscar Frith-Macdonald, one of our distinguished developers at DF. Oscar has studied the use of Machine Learning to facilitate image content searching for the UK National Archives. Who better to educate our team on the inner workings of artificial intelligence? His explanations are so excellent that we’ve invited him to write for the esteemed Weetbicks blog.

Take it away Oscar...

FileMaker — Now With AI!

For a while now, you have been able to access AI through the insert from URL script step and standard API calls (see Cath’s blog article). But FM 2024 is integrating a lot more of this directly into FileMaker, making it easier than ever to use AI in your solutions.

Now, there is a lot of great information on the cool things AI can do and also on the rather funny times it gets things wrong. However, we feel there is a lack of information on exactly what AI is. So this article is to help give you an understanding of how AI works, and this should help prevent you from falling into some of the pitfalls:

We will cover what AI is, some of the dangers of using AI, and how it can be useful. Hopefully, this will give you an idea of how best to use the new features safely.

So what is AI?

The best way to think of AI is as the outer shell of a Russian doll made up of deep learning (deep neural networks), artificial neural networks, and machine learning.

To break it down:

1. Machine learning (ML) is a subset of AI that involves training algorithms to learn from and make predictions based on data.
2. Artificial neural networks (ANN) are a subset of ML that mimic the human brain's interconnected neuronal structure to process information in layers.
3. Deep learning is a further specialisation within ANNs, utilising multiple layers (deep networks) to analyse complex patterns in large datasets.

In essence, each layer builds on the previous one, advancing the capabilities and complexity of AI.

Artificial Neural Networks

An artificial neural network (ANN) is an interconnected group of nodes, similar to the vast network of neurones in a brain.

Here, each circular node represents an artificial neurone, and an arrow represents a connection from the output of one artificial neurone to the input of another.

An ANN is composed of layers of interconnected nodes, or neurones.

These layers are typically organised as follows:

1. Input Layer: This layer receives the raw data. Each neurone in the input layer represents a feature or attribute of the data. For instance, in image recognition, each neurone might represent a pixel value.
2. Hidden Layers: These layers process the input data through a series of transformations. An ANN can have one or more hidden layers, where each layer consists of neurones that apply a mathematical function to the inputs and pass the results to the next layer. The term "hidden" refers to the fact that these layers do not directly interact with the external environment; they only communicate with other layers within the network.
3. Output Layer: This layer produces the final output of the network. The output can be a single value (e.g., a prediction or classification) or a set of values (e.g., probabilities for different classes).

Each connection between nodes has a “weight” to it, which is adjusted during the training process to minimise errors in the output. This process is similar to learning in biological brains, where synaptic strengths change to improve performance.

Deep Neural Networks

Deep Neural Networks (DNN) are a specific type of artificial neural network, with multiple layers in between the input and output layers.

A simple ANN may have just one hidden layer (shown in the diagram above), and a DNN adds many more of those hidden layers.

These additional layers enable DNNs to model intricate patterns and representations, making them highly effective for a wide range of tasks, particularly those involving large and complex datasets.

So what's happening in these hidden layers?

Each hidden layer transforms the input features into increasingly higher-level abstractions. Initially, the network might extract simple features (e.g., edges in an image). As the data passes through more layers, the network can recognise more complex patterns and structures (e.g., shapes and objects).

For instance, in image processing:

1. First Hidden Layer: Might detect edges and basic textures.
2. Intermediate Hidden Layers: Might detect patterns and shapes, such as corners or circles.
3. Deeper Hidden Layers: They might detect more complex structures, such as faces or objects.

Machine Learning

Machine learning (ML) focuses on developing algorithms that enable computers to learn from and make predictions or decisions based on data.

Unlike traditional programming, where explicit instructions are given to perform a task, machine learning allows systems to learn and improve from experience without being explicitly programmed.

Machine learning consists of three main points:

1. A Decision Process: Machine learning algorithms predict or classify based on input data, whether labelled or unlabelled, to identify patterns within the data.
2. An Error Function: This function assesses the model's predictions by comparing them to known examples, thereby determining the model's accuracy.
3. A Model Optimisation Process: To improve the model's fit to the training data, the algorithm adjusts weights to minimise the difference between actual examples and predictions. This iterative process of evaluation and optimisation continues until the desired accuracy is achieved.

Data Vectorisation

Data vectorisation is a key component of many AI models, though not all, which is why it wasn't included in the earlier Russian doll analogy.

Data vectorisation involves transforming raw data into a numerical format suitable for machine learning algorithms. Essentially, it converts data into numerical vectors, which are arrays of numerical values representing various features or attributes of the data. This numerical representation allows qualitative information, such as meaning to be quantified and compared to find similarities.

When data is vectorised, the numerical values generated are referred to as "dimensions." The more dimensions a vector has, the more accurately it can represent the data and facilitate comparisons.

Consider a simple example where the only two words in your vocabulary are "door" and "wall." In this case, the vectors might only have two dimensions:

1. "door" : [1, 0]
2. "wall"  :  [0, 1]

In other words, this limited vocabulary would be represented as [0, 0]. Comparing "door" and "wall," we see they are distinctly different in this simple two-dimensional space. If we introduce the word "window," its vector might look like this: "window": [0.6, 0]. Comparing "window" to "door" and "wall," we can see that "window" is more similar to "door" than "wall."

In reality, vectors typically have many dimensions, representing various attributes or features. These dimensions can be imagined in a virtual space, where the vector indicates a position within this space.

Cosine similarity is then used to determine how close two vectors are within this multi-dimensional space. By measuring the cosine of the angle between two vectors, we can quantify their similarity, ranging from -1 (completely opposite) to 1 (identical).

So, really, what is AI?

Now that you have an understanding of the different components that make up AI, let’s dive deeper into what AI truly is.

Knowing that a house is made of bricks and cement doesn't fully convey what a house is; it’s the same with AI.

When you hear the term ‘AI,’ you might think of iconic characters like HAL 9000 from 2001: A Space Odyssey or Data from Star Trek. These examples represent machines with the capacity for reason and autonomy, known as artificial general intelligence (AGI). However, this is not what most modern AI refers to. Today, when we talk about AI, we're actually discussing Narrow Artificial Intelligence (NAI).

Narrow Artificial Intelligence (NAI)

NAI is essentially an excellent piece of prediction software designed to perform specific tasks. Unlike AGI, which has the ability to understand, learn, and apply knowledge across a wide range of tasks at a human level, NAI is specialised. It's important to remember that it will only be an excellent piece of prediction software for the inputs it was trained for.

Here are some key points to understand NAI better:

• Task-Specific: NAI is designed to perform a specific task or a set of closely related tasks. For instance, a language model like ChatGPT is trained to process and generate text but cannot create images.
• Data-Driven: The effectiveness of NAI depends heavily on the quality and quantity of data it has been trained on. It makes predictions and decisions based on patterns found in this data.
• Prediction-Oriented: At its core, NAI is about making accurate predictions. Whether it’s forecasting sales, recognising speech, or recommending products, NAI’s strength lies in its ability to predict outcomes based on input data.

The importance of training data

The performance of NAI systems is tightly linked to the data they are trained on:

• Relevance: The training data must be relevant to the task. A language model trained on diverse and extensive text data will perform well in text-related tasks.
• Volume: Larger datasets typically result in better performance, as they provide more examples for the model to learn from.
• Quality: High-quality data, which is accurate and well-labelled, leads to better predictions.

There are three main forms of training, and all are used within the AI space.

1. Supervised Learning:

Supervised learning plays a significant role in training models like GPT. During the initial pre-training phase, the model learns from a large corpus of text data that includes pairs of input (e.g., text snippets) and output (e.g., the next word or sequence of words). This helps the model understand language patterns, grammar, and context. Key points include:

• Pre-training Phase: The model is trained on vast amounts of text data from the internet in a supervised manner to predict the next word in a sentence.
• Fine-tuning: After the initial pre-training, the model can be fine-tuned on specific datasets with labelled data for particular tasks, further enhancing its performance.

2. Unsupervised Learning:

While the initial phase of pre-training can be considered supervised learning due to the nature of the task (predicting the next word), it also has aspects of unsupervised learning:

• Learning from Unlabelled Data: The massive text data used in pre-training is not explicitly labelled with specific tasks but consists of raw text. The model learns patterns, structures, and relationships in this data without specific labels guiding the learning process.

3. Reinforcement Learning:

Reinforcement learning has been notably applied in models like ChatGPT for specific aspects of training, such as fine-tuning the model's responses to be more aligned with human preferences.

• Reinforcement Learning from Human Feedback (RLHF): In the case of ChatGPT, after the initial supervised learning phase, the model undergoes reinforcement learning, where human feedback is used to reward desirable responses and penalise undesirable ones. This process involves:
• Human Labellers: Provide feedback on model outputs, ranking them based on quality, which is then used to further fine-tune the model.
• Policy Optimisation: Adjusting the model's parameters to maximise the reward signals obtained from human feedback.

While NAI is powerful, it comes with limitations:

• Lack of Generalisation: NAI cannot generalise knowledge beyond its training. A language model won't understand visual data unless it has been specifically trained to do so.
• Dependence on Data: NAI’s accuracy is limited by the data it was trained on. If the data is biassed or incomplete, the AI’s predictions will reflect these issues.
• Absence of Understanding: NAI does not possess true understanding or consciousness. It operates purely based on learned patterns, without any comprehension of the content.

NAI and GPT

It's pretty difficult to talk about AI and not mention OpenAI and GPT at all, so how does GPT fit into the world of NAI? Let's start with what GPT stands for:

Generative Pre-trained Transformer

• Generative: This refers to the model's ability to generate text or other forms of output. 
• Pre-trained: Before fine-tuning for specific tasks, the model undergoes extensive training on a diverse dataset to learn language patterns, grammar, and general knowledge. This pre-training allows the model to perform well on various tasks with minimal additional training.
• Transformer: This is the underlying architecture of the model. Transformers are a type of neural network designed to handle sequential data and have become the standard for natural language processing tasks due to their efficiency and effectiveness in understanding context and relationships within the data.

GPT is actually an excellent example of NAI, as it's very obvious that none of the GPT models can do everything. They have had to be narrowed down to specific tasks such as ChatGPT, ImageGPT, etc. As their names suggest, they have similar base GPT models but have then been trained for specific tasks.

These different models can work in conjunction with each other to provide a “wider” AI experience. This is why ImageGPT can generate an image from a text input.

Example Workflow

1. User Input: The user provides a text prompt, e.g., "a two-story pink house shaped like a shoe with a small garden in front."

2. Text Processing: A language model (like ChatGPT) processes the text to ensure it is clear and extract essential elements, e.g., "two-story pink house," "shoe shape," "small garden in front."

3. Image Generation: The processed text is fed into a text-to-image model (like ImageGPT), which then generates an image based on the description.

AI Considerations and Mitigations

While AI is a powerful tool, there are important considerations and potential pitfalls to be aware of. Understanding these can help you better navigate the challenges and maximise the benefits of AI.

Pitfalls and Issues:

• Bias in AI: AI systems learn from the data they are trained on. If this data contains biases, the AI will likely reproduce and even amplify these biases. For instance, an AI trained on biased hiring data might favour certain demographics over others. It's crucial to ensure that training data is as unbiased and representative as possible. One of Google's issues with the AI summary was pulling information from places like Reddit and presenting it as fact.
• AI Hallucinations: AI models, especially those dealing with language, can sometimes generate outputs that seem plausible but are completely incorrect or fabricated. This is known as AI hallucinations and occurs because the AI predicts text based on patterns it has seen, even if those patterns lead to false information.
• Overfitting: An AI model might perform exceptionally well on training data but fail to generalise to new, unseen data. This issue, known as overfitting, can lead to poor performance in real-world applications.
• Data Privacy: AI systems often require large amounts of data, which can raise concerns about data privacy and security. It's important to handle personal and sensitive information responsibly, ensuring compliance with data protection regulations.
• Complexity and visibility: Many AI models, particularly deep learning models, are complex and operate as "black boxes," making it difficult to understand how they arrive at specific decisions. This lack of transparency can be problematic, especially in critical applications like healthcare or finance.

AI Hallucinations and Incorrect Results

Hallucinations are probably one of the biggest issues with AI, especially when the AI will provide an answer that seems plausible and will also present it as fact.

Hallucinations can happen for several reasons:

• Incomplete Data: The AI might fill in gaps with fabricated information if the training data is incomplete or lacks context.
• Pattern Recognition: The AI tries to predict the next piece of information based on patterns, which might lead to incorrect associations and results.
• Ambiguous Queries: Vague or poorly defined inputs can cause the AI to generate misleading or incorrect outputs.

To mitigate these issues, it’s important to:

• Validate Outputs: Always cross-check the AI’s outputs with reliable sources or experts in the field. N.B. As it is a well-known issue that these generative models can generate false information, many AI models have now been designed to return links to their source so that it is easier for a user to verify the information the model has returned.
• Provide Clear Context: Ensure that queries are specific and provide enough context for the AI to generate accurate responses.

Improving AI responses and ensuring correct information

To enhance the reliability and accuracy of AI responses, consider the following strategies:

1. Use High-Quality Data: Ensure the training data is high-quality, representative, and as free from biases as possible. Regularly update the data to keep the AI model current. Use your own existing data and examples when asking a question about AI; this will help to ensure it's only using information you consider to be true.

2. Implement Human Oversight: Use AI as a tool to speed up human decision-making rather than replacing it entirely. Always have a human look over the result to ensure it is correct.

3. Clarify Queries: When interacting with AI, especially language models, frame your queries clearly and precisely. The more specific the input, the more accurate the output is likely to be. Don't be afraid to send pages of information in your prompt to the AI, but be aware there is usually a cost for this.

4. Employ Ensemble Methods: Combine multiple models or methods to cross-verify results. Ensemble techniques can help improve accuracy and reduce the likelihood of errors. It's also important to use the correct model for your task.

5. Educate Users: Ensure that users understand the capabilities and limitations of AI. Providing training on how to interact with AI systems and interpret their outputs can significantly improve outcomes. One of the most important points is to ensure your users are aware that the AI could be wrong.