Other parts of this series:
Exploring AI part one: What on earth is AI anyway?
While there’s a huge amount of interest in both machine learning and AI, there isn’t a broad understanding of what they mean in practical terms. In part one of this series, we’ll demystify AI and machine learning and explain what they are and what they are not.
The adoption of AI is on the rise. However, without first understanding the AI landscape, the route to the implementation within business practices, and the benefits of doing so, can seem unclear. Whether you’ve made steps on your journey or you’re gearing up for change, it’s important to know where you are and where you aim to be. AI is a wide field and the aim here is to introduce the concepts of AI and how different flavours of it compare.
The AI/Machine Learning landscape
In its most general form, AI is simply any computer system that exhibits intelligent behaviour. It’s a subset of computer programming in general, though there are many computer systems that you wouldn’t really describe as being intelligent – a database for instance, arguably, does not exhibit “intelligence”, it just stores and retrieves data. But, what constitutes “intelligence” is a grey area and the boundary between AI and non-AI systems isn’t a sharp one.
Expert Systems: Doing as they’re told
One of the oldest forms of AI is the expert system: the collection and encoding of human based knowledge as a set of rules. A typical example of this “do as you’re told” intelligence is:
IF Variable1 >= 0 then True.
Here, we program the exact logic, from the variable to examine to the values of importance for a specific outcome. The logic is generally developed by surveying how human experts respond to a situation and producing a complex set of rules that attempts to replicate their behaviour. Programmatically achieving the outcome is the first sign of AI, skewed heavily towards the ‘artificial’ side of the term.
This approach initially speeds up the process of decision making and allows it to scale. However, as business complexity grows, as does the complexity and number of rules that are needed. Today, documenting, maintaining and updating these models comes with a greater time cost and risk factor than ever before. Conversely, these systems allow for specificity, investigation and clear justification at the level of an individual action; this can be seen as a benefit, if not a necessity, in heavily regulated industries.
Machine Learning: Learning what to do
Machine Learning (ML) is the application of algorithms that learn by themselves through experience. Common machine learning algorithms include decision trees, which derive a series of “if then else” rules automatically (and so in a sense are like automated expert systems) - and regression models which essentially identify lines of best fit through the data.
Let’s clarify exactly what this means. ML algorithms rely on training data, either sourced from real life observation or produced through a simulation. The features in the input data go through a process that aims to model an outcome, for example a decision tree, and outcomes are observed. ML is achieved through the measure of error in the outcomes and the gradual adjustment of the modelling process to reduce this error. The ability for these programs to learn from experience is the most important distinction between machine learning and general AI – the models are learned from data rather than hand-crafted by a human.
To contrast with expert systems, instead of an industry expert deciding the logic, ML finds the variables and values that are of greatest importance and which can be used to most accurately predict the outcome variable. It analyses the relationships between the variables in the data and uses this to produce a model in effect playing the same role in this process as the industry expert does in an expert system. The model is then used to make decisions.
Moving from an expert system to an ML model we see another level of efficiency that reduces the time taken to build complex models, compared to using industry expertise. However, there are a number of ML approaches that are possible and as such we need ML experts to select the correct ML path to pursue. Achieving good results also sometimes requires careful and creative data preparation to produce good transformations of the raw input data that are useful in the model – a process known as “feature engineering”. Again, this requires manual effort and expertise – though still considerably less than manual review or even creating an expert system.
Though ML can be applied in a broad range of situations using a wide range of techniques, in practice businesses tend to rely on only a small subset of them. A very common example being logistic regression, which tends to be more intuitive and allow for quicker investigation than other methods. This maintains the ability for businesses to demonstrate control and understanding of their model albeit somewhat reducing the potential power of the model. As the increase in predictive power that more complex ML models can offer becomes apparent and our ability to examine and explain them improves, we see a shift towards the exploration of advanced solutions.
Deep Learning: Learning what to learn
Deep learning is a type of ML that is based on using Deep Neural Networks (DNNs). We call them Neural Networks because they are loosely based on how neural connections in the human brain work. While neural networks have been around for many years, they have really come to prominence in recent years because of some of the valuable, high-profile use cases that they support. Self-driving cars, AlphaGo (now hands down the world’s best Go player), smart assistants and many more innovative new technologies are all powered by deep learning. It has been applied across a wide range of tasks to generate state of the art results and is enabling a wide range of fascinating and powerful applications. It’s almost single-handedly driving all the current hype around AI.
DNNs can be expressed as an architecture of artificial “neurons” that receive some inputs and generate an output based on them, which is then fed in as inputs to other neurons. These neurons are typically arranged in a series of layers, and the “deep” part of the name alludes to there being many such layers involved. Each neuron is relatively simple, though, due to the number of interactions between them, DNNs have considerable capacity to express complex relationships in the data and are able to identify them automatically, meaning we increase efficiency because there is no need for manual feature engineering to get good results – they can do all that work for themselves, in effect learning what’s important to our outcome, or ‘learning what to learn’.
The models can be very complex and as their size and complexity grows, so too do the computation requirements. The complexity also means models are hard to understand and have the opportunity to behave less intuitively when we examine their output. Enhancing our understanding of deep learning outputs is an area of active research, though when we do take time to thoroughly examine these complex interactions, we see how deep learning is allowing us to discover relationships in our data that were previously hidden and too complex to for our models to express, as such we gain new insight in to our world.
With AI and machine learning set to become such a prominent feature in the future of business, whatever your background, having a clear understanding of the terminology and concepts associated with AI and ML is key to navigating through all of the hype associated with it and understanding how and why it’s able to help with a range of tasks. With an understanding of what AI and machine learning is, you’re better equipped to understand how it can be implemented within your business and the benefits of doing so.