Artificial Intelligence is a field which seems to be evolving at breath-taking pace, and looks set to be one of the most significant technological advances of the 21st century. As such it is a topic which we will return to time and again over the course of this programme. But what exactly is ‘AI’, and how does it work? Dr Ken Amor investigates.
Artificial Intelligence (or simply AI) is becoming embedded into many aspects of human society. When you type an entry in Google search, it uses an AI system to return the websites that most closely match your query. In doing so, it has to interpret the meaning of your query by examining the context of surrounding words, while correcting for spelling mistakes. ‘Change’ for example has many different meanings in the English language, such as replace, exchange, adjust, and so on. Google search also considers your historical and recent searches, to return the most relevant webpages first.
So just what is AI? It’s often defined as an intelligent machine that can think and act rationally and is capable of problem-solving and decision-making, just like the human mind. AI is a general term to encompass many different types of intelligent machines that are constructed to perform a specific task or application. Typically, “intelligent” computer programs lie at the heart of any AI system. There are six main objectives in AI: reasoning, knowledge perception, planning, learning, natural language processing and perception. (A seventh goal is to move and manipulate objects, but this applies to the combined disciplines of computing and robotics.)
AI consists of many different components, such as machine-learning and artificial neural networks, using methods taken from statistics, probability, economics, computer science, psychology, linguistics and philosophy, although not all AI systems use all these elements. A common feature of AI is access to large amounts of information called datasets. Advanced AI systems can interpret and use raw information such as speech or satellite data without the need for humans to classify or organize it in a structured database. AI uses algorithms, which are a set of specific instructions to accomplish some task, such as perform a calculation, or make a pre-defined reasoning decision, such as If A is True then do x, If B is True then do y etc. Conventional computer programs also use algorithms, but in artificial intelligence, these algorithms can be trained to improve automatically using real data, and without a human programmer having to change lines of code.
A simple example of a machine learning algorithm is an email spam filter whereby the program classifies incoming email into spam and not spam. Depending upon your actions: delete without reading, read and then delete immediately, keep etc., the filter learns to recognize incoming emails you regard as spam. One approach is the use of artificial neural networks that aim to mimic biological neural networks such as found in the human brain. Inputs into the system can be processed by a number of different algorithms connected by different pathways (neurons). A particular algorithmic pathway is given a weighting, depending upon the success or otherwise in achieving the goal, in this instance deciding whether an email is spam or not. Over time the best solution is given by the pathway with the highest weighting. In this way AI systems can respond to new data without having a human computer programmer change the program.
How does this new technology differ from conventional computer programs? To illustrate the difference let’s take as an example a recipe to make pancakes. A recipe is a set of instructions which, if followed correctly, results in a delicious meal. If the recipe only lists the traditional lemon juice and sugar topping, what happens if you don’t have any lemons in the kitchen? You might look through the kitchen cupboards for an alternative. It is human nature to be inquisitive and explore alternative ingredients and combinations of foodstuffs that you may not have previously tried before.
A computer program also follows a set “recipe” as a list of instructions. Options and conditions can be hard-coded into the program, but in this comparison, the computer program stops because it is unable to complete its list of instructions – no lemons. However, an AI system would also explore other alternatives. AI systems have access to much more data and would “know” the contents of your store cupboard, perhaps by analysing the purchases you make at the supermarket. It may classify lemons as a citrus fruit and if no lemons were available suggest an alternative such as oranges. The AI system may also note that the traditional topping consists of a combination of sugar and fruit, and therefore suggest a sweet, tangy alternative such as jam. This exploration mimics the human mind, its reasoning, knowledge and perception of the world. Not all combinations of foodstuffs will make a successful pancake topping but the AI system would learn from its mistakes just as we do, in this instance perhaps by a using a feedback form.
Dr Ken Amor is the Access & Outreach STEM Associate at St John’s College