Artificial Intelligence and Machine Learning – what are the differences?

Artificial intelligence is one of the hottest topics of the last few years, both within the IT industry and with the general public. What makes AI different from other buzzword-based trends is that it is an important part of Computer Science, which has existed since the middle of the previous century. Back then, AI was mostly related to scientific research and it rarely came up in real business applications and everyday discussions. However, the last decade brought a significant change and we can safely say that AI is present in many different aspects of our lives, even if we don’t notice it at once. 

Contrary to many other popular IT trends, such as Blockchain or Augmented Reality, AI is an extremely broad, umbrella-like term, covering different algorithms and concepts, since the definition of AI is very wide and imprecise. In order to approach the AI definition, you need to think about the common denominator – what brings all the “intelligent” algorithms together?

While the common understanding of AI has been changing over the years, this seems like a good approach. It is usually assumed that AI is used for fulfilling all the activities that are considered a sign of human intelligence. And I think we can all agree that this definition is very broad, but at the same time it’s a very good one, because it does not focus on technicalities.

But even this definition has flaws. Although in some cases, like with computer programs playing chess, computers are much better than even super grandmasters, such as GM Magnus Carlsen himself, other tasks that might seem trivial for us are still a great challenge for AI, e.g. recognizing different species of birds or dog breeds. While these problems have gained some noticeable improvement just in a few recent years, they were extremely difficult to solve for a long time.

Typical problems solved by different AI methods are: image/sound/video recognition, natural language processing (NLP), generating music, images and – especially recently – texts (ChatGPT, anyone?), especially in terms of translations or content marketing. Another important application – although it might not be as popular as the previous ones – are intelligent decision systems and fuzzy logic, applied both in the industrial environment and computer games.

While we discussed the Artificial Intelligence definition itself, nowadays it is almost impossible to see it without its important counterpart – ML. Machine learning is *not* equivalent to artificial intelligence, although its relationship with AI is important and rather simple. ML is just part of AI that concentrates on some of the tasks, related especially to creating models that improve their quality iteratively based on presented data. This definition might sound vague, so let’s jump straight to the examples. One of the simplest ones I can present – the one that is close to my heart, as it was the topic of my master thesis and it brought me to the scientific world over 10 years ago – is speech emotion recognition. I used a well-known German database called EMO-DB. This database contained hundreds of sentences, where every sentence was spoken with a single emotion (e.g. happy, sad, bored, angry, etc.). All these sentences were used to “teach” a machine learning classifier so later it could detect emotions in sentences that had never been presented before. 

The tricky question is – what does it mean to “present”, “show” or “teach” a machine learning model when we have, for example, a spoken sentence? Let’s say I had a German sentence “Ich habe eine Kartoffel” (I have a potato), spoken angrily. The classic approach I applied over 10 years ago consisted of two main steps:

• First, I used the sound signal of this sentence to calculate a few important characteristics, for instance, what the energy of the signal was, what the pitch (high/low frequency) was and how much it changed over the sentence, as well as a few more complex ones. In case of an angry sentence, I would probably obtain high energy and quite a significant change in frequency. When you speak angrily, you usually change the voice subconsciously, in order to emphasize your message.
• Then I passed all the above parameters (basically a bunch of numbers) as the input for the classifier (a neural network), at the same time telling the classifier that this particular set of parameters should be associated with ‘an angry sentence.’

Of course, to achieve a properly taught machine learning classifier, I needed to perform this operation many times for plenty of different sentences. In the end, after the classifier finished its learning procedure, I could speak a sentence to the microphone and the program was able to tell me what my emotional state was.

I wanted to introduce some basic terms in this first – and certainly not the last – article about AI, but let’s move forward and tackle another controversial AI-related issue. What is currently the biggest challenge when delivering AI-powered IT projects?

This question is of course very vague and certainly there can be many various opinions, but if I were to choose one, I would say that people often forget that AI-powered projects have inherent risk, much higher than in case of “regular” IT projects, even those based on custom software development.

Although every IT project might have some obvious risks, mostly related to mismatch between business needs of the client and what the IT team delivers, using AI in a project belongs typically to a research and development (R&D) domain. When you decide to include some AI features, especially based on custom datasets (e.g. specific to a particular organization), there is always a risk that your business assumptions, your working hypotheses will not be confirmed once you have completed your AI features. It is not just like with a web application development process. Let’s say you need to implement a list of all invoices in a web application. The risk lies mostly within the estimating process itself – this task might take more or less time – but you don’t need to worry about being able to implement it at all.

To give you another example, there are many computer science problems that have been solved accurately by algorithms that we describe as deterministic (in contrast to metaheuristics, on the topic of which Kasia wrote an article recently). For example, any time we want to sort a list of numbers or texts, we can be 100% sure that, eventually, we will receive a sorted data set, from A to Z. AI solutions (and especially ML algorithms) are not as precise and deterministic, so we cannot be sure if, for example, we just created a perfect classifier that distinguishes between dogs and cats (a classic basic problem of machine learning classification). In some cases it might be even impossible to say, for instance, in the speech emotion recognition, perception of emotions can be different for different people and in different languages, so sometimes it is quite difficult to assess what *should* be the correct result of AI’s work. Another problem is the matter of performance. Even an algorithm that works correctly – e.g. recognizes human faces in the presented images – might be not sufficient, if it works slowly and cannot be used for real-time face detection. 

The last issue I would like to address now is the existence of strong AI, also called Artificial General Intelligence. In short, constant progress in AI achievements brings the important question – how close are we to creating AI that will be self-conscious and will essentially become equal to what we, humans, can achieve? It’s difficult to summarize quickly, but I will try nonetheless. In my opinion, this cannot be achieved in any foreseeable future. Even though the progress in AI, especially within the field of Deep Learning (which I will discuss in the next article), is truly spectacular, it still does not touch in any way on self-consciousness or ability to learn completely new skills. The new models we see every year in AI could be compared to inventing new, faster ways of moving on the ground – running, riding a bicycle, driving a car. But in order to achieve AGI, we would need to learn how to fly to the skies & outer space, to other galaxies and beyond. So the difference is humongous.

Does this mean we shouldn’t worry about using AI? Not at all. The biggest fear we should care about now is the context in which we put all the AI solutions. Any place, any system where we let AI make decisions, for example in medical diagnosis, in autonomous vehicles or especially in the military, poses a certain risk. The real problem is whom to blame if any of the intelligent systems go wrong. The neural network model creator? The dataset manager? The solution producer? Or maybe the end user who overlooks the AI behavior? It is certainly a challenge that should be discussed more in-depth.

This is the real issue we should focus on now, but the AGI problem is not something we should completely forget about. One of the biggest fears of AGI is that it might result in creating the Singularity – once AGI becomes as intelligent as we are, it might use its computational power to constantly improve its quality, leading to incredibly fast development and becoming the higher entity that will ultimately control us. But that’s definitely a story for another time.