Neural Networks

How does the brain use this process to respond to something like heat?

For instance, if you accidentally touch an activated electric hob, the neurons sensing temperature signal to the neurons controlling your hand that the hob is too hot, prompting you to move your hand away. How quickly you need to move depends on exactly how hot the hob is:

‍

‍

These thresholds are typically learned through trial and error. A child is more likely to reach toward a hot hob than an adult, because they haven't yet learned it could hurt them. If they do touch it, their neurons "remember" that this is something to avoid in future.

How do Artificial Neural Networks (ANNs) work?

Artificial neural networks (ANNs) work in a similar way to our bodies. They're built from artificial neurons (or nodes) arranged in layers that work together to solve a problem.

• Information from the environment enters the ANN through the input layer, which processes and analyses it (e.g., working out how hot the hob is) before passing it to the next layer.
• The middle layers determine the action that should be taken (e.g., move hand away).
• The output layer produces the final result (e.g., the hand moves).

Just as in the human brain, the meaning of a given piece of information, and what to do about it, is learned through trial and error. The neural network "looks at" lots of data, guesses what to do, assesses whether its guess was right or wrong, and refines its actions over time.

ANNs used in healthcare aren't assessing whether a hob is hot, of course. They're mainly used for classification tasks, such as examining a medical scan and deciding whether it contains a feature of interest. But the underlying mechanism, taking in information, categorising it (e.g., deciding whether a tumour is present), and assigning a label, works the same way.

What are the structural components of an ANN?

ANNs are designed for pattern recognition and form the building blocks of many of today's most successful deep learning models, including large language models like ChatGPT.

They're made up of node layers: an input layer, one or more hidden layers, and an output layer. Each node processes, analyses, and categorises information passed to it from the nodes in the preceding layer, then passes on what it has "learned" to the nodes in the following layer.

These layers are hierarchical:

• Outer layers handle more general calculations and analytical tasks
• Inner (hidden) layers handle more specific or complex tasks, and do most of the heavy lifting
• The input layer is where the information being assessed (e.g., an image) enters the network
• The output layer gives the final result of the calculation (e.g., a label for the image)

How does an ANN decide what information matters?

Each layer of nodes, and each individual node, analyses incoming information (in the form of variables) to determine how important it is to the question being asked (e.g., "is this an image of a cat or a dog?"). Importance is denoted by weights, variables deemed more important are assigned higher weights.

All inputs are multiplied by their respective weights and added together. If the result exceeds a certain threshold, the node "activates" and passes the data on to the next layer. This is why most neural networks are described as feedforward networks: the output of one layer becomes the input for the next.

The variables being weighted get progressively more specific with each layer. Outer layers might focus on high-level features, like the edges of shapes in an image, while inner layers examine increasingly specific details (e.g., the eyes, ears, or tail of an animal), before ultimately classifying the image as a dog or a cat.

What are the main types of neural networks?

The right type of neural network largely depends on whether the task is one of classification or prediction. The most commonly used types include:

• Deep neural networks
• Convolutional neural networks (most often used for image analysis)
• Recurrent neural networks
• Generative adversarial networks

What are the advantages and disadvantages of ANNs?

Different types of ANNs come with their own advantages and disadvantages, which vary slightly depending on the specific model used. These need to be carefully weighed during model development.

Advantages:

• Ability to detect complex, nonlinear relationships between variables
• Ability to cope with noisy or missing information
• Ability to generalise to similar, unseen data
• Ability to learn quickly and efficiently

Disadvantages:

• Unpredictable results (random weight assignment can produce multiple different solutions to the same problem)
• Limited ability to provide insight into the causality of identified relationships
• Inscrutability (the complexity of hidden layers makes it difficult to explain how outputs are generated)

‍