AI in our world17/07/2019
“Will AI replace humans?”, it’s a question a lot of persons seem to be preoccupied with, but what’s the reality? There are two approaches to the subject of AI: the probabilistic one and the machine learning one.
First of all, the most conventional one, which has been in place for many years, the probabilistic and logistic one. We know what AI is doing as everything has been programmed by someone. Computers therefore have supposedly no autonomy and can only execute what their programmer wrote them to do.
Then, the newest approach is machine learning. Though the concepts behind it were proposed decades ago, the required computational power has only been made available in recent years.
Compared to what we know from standard AI algorithms, machine learning acts more like a black box. In fact, we create some algorithms and then allow the AI to train on data. The AI is then capable of reproducing patterns that it has learned and sometimes it even discovers relationships we were not even aware of. This is what makes people afraid of Artificial Intelligence; machines are capable to carry out some kind of thinking and are capable of doing things “like human”. This terminology is quite important as they mimic human reactions to problems based on what other people have instructed them. However, the human brain is very complex and there are more parameters influencing our reactions than we might consider, meaning the computer may not have all the required data, it will just guess. The quantity of reactions and actions is also very large and the way they work is as varied as the number of possibilities. The machines are then only able to do one thing they have been trained for, meaning we will need to train one machine for each type of reaction/action which is not possible in practice.
The possibilities with machine learning are, as we have seen, very large and the tasks that it can handle, can easily be divided into three main families: classification, regression and clustering. Classification tasks aim to classify inputs into categories, for example deciding whether an animal is a cat or a dog based on its properties. Regression tasks aim to find a value for a certain input; this is useful when we do not know the formula underlying a phenomenon, the AI can find an estimator for this formula and predict a value, like predicting the fuel consumption of a vehicle based on its technical data. The last category is a clustering task or anomaly detection task; the goal is to detect anomalies in huge data sets, like detecting hacking attempts in a company’s internet workflow.
But how does it work in practice? We often hear about neural networks but they are not the only solution. Modern machine learning tools are a combination of different algorithms exploiting different strong and weak points of each technique to enhance the results. Alongside the neural network, we can cite decision trees, regression algorithms, Bayes estimators, etc… The quantity of available algorithms is very significant and tuning of those algorithms is required.
Due to this quantity of knowledge and the specificity required for each task, the science fiction vision of the world is then no longer something which could occur in the next few years and according to certain specialists in the domain, will never occur.