How AI works

• Artificial Intelligence is a broad field that tries to copy human intelligence with computers.
• AI does not try to copy the human processes, but the results. The way to achieve those results is actually very different from humans.

• There are many approaches and many different techniques in the field of AI, and they all have one thing in common: the task must be narrowed down and framed in terms of something computers are good at. Computers excel at doing mathematical calculations, they can perform them much faster than humans, and without fatigue.
•  Let’s take chess as an example, computers have been able to beat humans at chess since the 90’s. They do not understand chess tactics, but they have the computational power to brute-force their way to a win.
• AI researchers are not interested in how human minds work when playing chess, but they try to replicate its outputs using clever algorithms and mathematical modelling.

• The strategy they use to make computers choose their moves is known as a ‘game tree’.

• Starting from an initial position, the computer lists every legal move that it can make and, for each of those legal moves, it considers every possible countermove that the opponent could make, and so on. The graphical representation of the possibilities is a diagram that resembles a tree, with each move creating new ‘branches’ of possible countermoves.

• The computer searches through the branches, ‘pruning’ those that have minimum winning outcomes and keeping the ones with minimum losing outcomes, in an attempt to find the best path to a win. When the branch with the higher winning chances is identified, it chooses the move that is at the beginning of that branch.
• Chess is a very complex game; each move creates multiple branches that quickly increase the number of calculations that a computer would need to make to solve every playable position. In fact, it is not possible to calculate all the possible moves and ‘solve’ chess, at least not with the current technology.

• Computers are therefore programmed to calculate only small portions of the chess game tree. At the same time, they compare the position with large databases of real-life games played by grandmasters to try to predict the outcome. The result is not perfect, but it is good enough to beat human players.

• Game trees are not the only approach to developing AI algorithms, you can learn more about other techniques in our quick guides: Machine Learning and Neural Networks, Generative AI. Researchers are finding new ways to simulate intelligence in computers every day, but they all require:
   
  1. that the computers can do calculations with the data they are analysing, and
      
  2. that specific criteria are defined for deciding which factors should be minimized or maximized.

References

• Gerrish, S. (2018). How smart machines think. MIT Press.
• Boden, M. A. (2018). Artificial intelligence: A very short introduction. Oxford University Press.
• Hsu, Feng-Hsiung (2002). Behind Deep Blue: Building the Computer that Defeated the World Chess Champion (1st ed.). Princeton University Press.

Machine Learning and Neural Networks

• When an AI model learns from data instead of being manually programmed for every task, that is known as Machine Learning (ML). It’s like teaching a child by showing examples instead of giving step-by-step instructions.

• Computers can do calculations quickly and precisely, combined with clever programming, those traits make them good at recognizing patterns.
• In the ML field, programmers give computers many examples and program them to detect patterns and ‘learn’ to identify specific characteristics in the data. For example, the programmers may give an AI model millions of photos of cats and let it find the patterns that all the photos have in common so that the computer can identify cats in new images.
• The 'clever programming' used for ML is a type of algorithm called neural networks. They consist of layers of interconnected simple nodes that process information step by step. This arrangement of simple units that working together can handle complex problems was inspired by the human brain, which is why the nodes are also called neurons.

• Each node receives inputs, applies a mathematical operation, and passes the result to the next network layer. For instance, one node might be specifically trying to find a furry contour in the image, another one might be specifically looking for two triangles that could be the ears, and so on. Each layer refines the information as it passes through until an output layer is reached where a result is produced for the user. The result is expressed as the result of those operations in probability terms, such as, a 73% chance that this picture contains a cat.

• There can be many nodes in each layer and many layers in a model; the configuration is specific to each application and purpose. Evidence suggests that the most efficient way to represent complex functions is by going deeper instead of wider, that is, using more layers with fewer nodes per layer. These architectures are known as deep neural networks, and the state-of-the-art models have reached thousands of layers in recent years.
• With this elaborate structure, machine learning and neural networks can handle tasks that are too complex or time-consuming for humans. But these methods are imperfect, they can make mistakes, especially if the training data is flawed, biased, or incomplete.
• Machine learning is a tool to help humans solve problems faster and more intelligently, but it needs careful guidance to ensure it works fairly and accurately.

References

• Samek, Wojciech, et al. "Explaining deep neural networks and beyond: A review of methods and applications." Proceedings of the IEEE 109.3 (2021): 247-278. doi:
  10.1109/JPROC.2021.3060483
• Gerrish, S. (2018). How smart machines think. MIT Press.
• Montufar, Guido F., et al. "On the number of linear regions of deep neural networks." Advances in neural information processing systems 27 (2014).
  arxiv.org/pdf/1402.1869

Black Box

• A common question is whether a black box exists in AI Models. The short answer is ‘no’, but a more accurate answer would be ‘it’s complicated’.
• The concept of a black box in AI refers to situations where the inner workings of an AI system cannot be understood or explained, sometimes even by its creators. This question is common when discussing complex models like neural networks. Example: A neural network might accurately classify an image as a cat but not provide an understandable explanation for how it made that decision. 
• Why is it a Concern?
  • Accountability: If AI systems make decisions that affect people (e.g., loan approvals, hiring, or medical diagnoses), it's critical to understand why they made those decisions.
  • Bias and Fairness: Without transparency, detecting and correcting biases in AI models is harder.
  • Trust: People and organisations are less likely to trust AI systems they cannot explain.
• Despite their biological inspiration, there is nothing mysterious about neural networks. They evaluate a series of mathematical formulas, and their operation has no randomness, uncertainty, or magic. The connections between the neurons dictate the results, and those connections are tuned by the training data used.

• The complications arise in understanding what the AI model does. It is not impossible to understand, but it may often be impractical to spend the amount of time required to do so, not to mention the knowledge level needed in the field.
• AI models, particularly deep learning models, rely on millions or even billions of interconnected parameters and layers to identify patterns and make predictions. This complexity makes it challenging for humans to trace how specific decisions are reached.
• A programmer can ‘look under the hood’ of the model and review every value present in every neuron, but the numbers would not make much sense because they represent relationships. The full explanation of why a neural network reached a specific output is encoded in the billions of diffuse connections between the nodes. The answer is not in the model; the answer is the model. 
• To point out the exact reasons for the decision, in a way that is coherent to humans, may not only require the analysis of the whole neural network but also the training data that was used to find the patterns that were identified by the network itself.

• Researchers are developing methods to make AI systems more interpretable. For instance, techniques like:
  • Model induction: Treating the system like a ‘lab rat’ and running tests and experiments to identify which factors most influenced a decision.
  • AI Probing: Using a second neural network that learns how individual neurons in a network respond to inputs and then creates a visualisation of the most important activity.
  • Simplified models: Creating interpretable approximations of complex models.
• In summary, while the complex nature of AI poses challenges, ongoing research aims to make these systems more transparent and accountable. In some applications, it is worth mentioning that the focus is on accuracy rather than interpretability (e.g., detecting diseases in medical imaging), so justifying the decisions is not strictly necessary. However, understanding the ‘black box’ becomes essential when AI affects human rights or high-stakes decisions.

References

• Gerrish, S. (2018). How smart machines think. MIT Press.
• Bleicher, A. (2017). Demystifying the Black Box that is AI. Scientific American. 
  https://www.scientificamerican.com/article/demystifying-the-black-box-that-is-ai/.
• Rudin, Cynthia, and Joanna Radin. "Why are we using black box models in AI when we don’t need to? A lesson from an explainable AI competition." Harvard Data Science Review 1.2 (2019): 1-9.
  https://doi.org/10.1162/99608f92.5a8a3a3d