The subject of Artificial Intelligence has collected more than its fair share of misinformation and hype. Terminology can become blurred and expressions incorrectly used. Read on for a clear guide to the critical points.
In AI most commonly we are working within a single topic. Whilst the subject might be thought complicated there is only so much to know and this can be learned and used in context: as an example London black cab drivers who ‘do the knowledge’. We refer to this as Artificial Narrow Intelligence. ANI tasks are typically described as those that a person could do with ‘less than a second of conscious thought’ ie you don’t have to think about it. That is because the knowledge is directly applicable to the subject. On the other hand, Artificial General Intelligence (AGI) is the capability to apply reasoning capacity in unfamiliar surroundings ie take what you know and use it out of context.
Consider the problem of a ‘self driving’ or more accurately a ‘computer driven’ car. In AGI the task would be very complex. It would involve taking existing known data not related to the activity of driving and finding a way of applying it. For example, cars are made of metal; metal is a hard object; when hit by hard objects people are hurt; avoid hitting people when driving the car. However, in the field of ANI where driving the car is the only concern, all that is necessary is to define how to respond to any given situation. When handling huge amounts of data is fast and cheap this can be done by mimicking what a human would have done in equivalent circumstances. So in ANI the task is simply to record many hours of people driving and mimic.
The most common form of AI is machine learning which comes in two flavours. In supervised learning a selection of past data is curated which shows that input A leads to output B and so supervised learning is also referred to as A to B mapping. Learning means to configure algorithms to execute the mapping. In unsupervised learning inputs A are provided along with the mapping. There are no examples of the outputs so results are unmoderated and can be unpredictable even nonsensical.
Supervised learning is at present more widely used. Examples are common in consumer internet such as ‘customers also bought’ recommendations.
Deep learning is so called because the patterns that drive activity sit below the surface. To uncover useful insights the inputs are processed in a number of intermediate steps to reach output B. Within these steps the elements of input A may be grouped, compared or processed several times and with different weightings to identify the appropriate output B. As an example ‘price’ as a factor in a buying decision might be used to consider cost, timing of purchase, and to measure desirability. The recombination of several factors in this way leads deep learning often also to be called a neural network.
AI, or more correctly ANI has advanced quickly because of three main factors – collection, storage and processing of data at a low cost. This is because AI needs lots of data! In fact, enough data such that the live inputs are sufficiently similar to training events to allow the algorithm to interpolate or extrapolate its way to a result. In practical terms this looks like a dataset with many measurements spread across many factors. The role of the AI is to process a set of measurements linking it with its known outcome. Then to do this repetitiously with many similar sets of data ie training. When a new set of data is presented the AI uses the training to predict the likely outcome.
Data volumes within consumer internet applications can be huge, but not every situation is like that. For example fault identification by visual recognition on a production line may have only a few examples. Next time we talk about the strategies to manage situations with low data volumes.