Artificial Intelligence is computer programs that predict an outcome of events or decisions like image recognition, human behaviour, or the future market value of something.

Machine Learning

If you compare AI to a motor vehicle then a Machine Learning (ML) model is the engine and Big Data the fuel.

As a vehicle’s engine can be combustion, steam, electric or rotary, the type of Machine Learning engines can include Deep Neural Networks (also referred to as “Deep Learning”), Convolutional Neural Networks (CNN), Random Forests, and Deep Reinforcement Learning (DRL).  Where engines can be 4, 6 or V8 cylinders, ML models use different algorithms depending on the application.   You don’t really need to know how they work, or how to build them, only what you want to achieve with them (see AI Insights Services – Machine Learning.  

Data Science

Where fuel has a chemical make-up (lead, ethanol, etc), data is made up of features where, like gasoline, some features affect performance more than others do. The big difference to vehicles is the amount of fuel AI engines need. Machine learning typically requires “Big Data” of 100,000s if not millions of examples (litres) with a more complex fuel mix than petroleum.  Data Science is the fuel science of finding, selecting,  cleaning and refining data to enable to best performance of ML predictions, see AI Insights Azure Cognitive Services – Data Insight.

Feature Engineering

As racing drivers need the right mix of fuel, so ML models require the right mix of feature data to achieve best results.  Depending on its use, a vehicle can often perform as well as on 95 octane as 98 octane, but at substantially less cost. Correspondingly, ML models can perform as well (same result) on less features with substantial less cost (processing time). If your AI is feeding customer service enquiries or monitoring a production line, reduced processing time can be the difference between a workable solution and a waste of time.

So we use “Feature Engineering” as the process of working out which features are driving outcomes, and dropping those that are just reflective of the causes.  

Model Training

Next, just as we tune vehicle engines, so we use weights and bias settings to tune accuracy of ML models. This tuning process is what we refer to as Training the model.  Training calculates the weights and biases through a reiterative process of estimating an outcome from historical data, comparing it to the actual outcome, adjusting the settings and then repeating to reduce the discrepancy. Obviously, this is a very tedious process, which is why AI is only a recent advent due to improvements in computer power and amounts of data available.

Model Accuracy & Performance

You would not select a delivery van without a test drive and working out its running cost. The equivalent in AI is measuring a models accuracy (test drive) and its performance (running cost).  How you assess suitability depends on your intended use.

Believe it or not, accuracy is not the only way you may want to assess an AI model.  Just like horse-power may not be the best measure of the van’s suitability. In addition to accuracy (% right) there is also Precision, for when a few wrong predictions could be critical e.g. the death penalty, so generally it’s better to use a combination of both known as the F1-Score. (see AI Insights Machine Learning Quality Assurance)

Model Re-Training

Finally, when in use your car needs periodic servicing and re-tuning.  Same with AI models. Although they don’t go out of tune, the world around them changes and unless you’re using Reinforcement Learning, they can’t adapt by themselves.

To Summarise the Analogy

Unfortunately, most information and services about AI on the internet relates to building your own engines. 

In business you wouldn’t build your own delivery vans, so at AI Insights we act as the vehicle dealer who assembles Microsoft’s “kit vehicles” to your requirements, puts them “on-road” and provides follow-up servicing (see Azure Cognitive Services).