What even is „AI“? 

AI (Artificial Intelligence), a term you hear a lot today. You may think this has something to do with R2D2 from Star Wars (Movie – 1977), the Terminator and Skynet from The Terminator (Movie – 1984). Or any other movie / book / video game where AI may receive consciousness. In reality it has nothing to do with this. Simply, there is no human like system and “intelligence”.
Generally, you can split the field of artificial intelligence into two subsidiaries: Rule-based Systems and Machine Learning. Let’s have a closer look:

Rule-based Systems (subsidiary of artificial intelligence):

These are human crafted and human defined Computer programs. Often used in Video Game AI, “Expert Systems” and Bots. You can imagine this like a giant if script:
If rain = 1 and sun = 0, then sales = 0.5. And so on. For every possible case and circumstance.
You may see the problem. You need to create a specific rule-set with the help of a human expert of the domain you are aiming for. Some cases are too complex and not solvable with this approach.

Machine Learning (subsidiary of artificial intelligence):

Unlike rule-based systems you don’t need human experts to program specific rules for this system. These systems are “learning” and defining their own “rules” from data. A lot of data. For example: Given a set of features (X) and one (or more) target/s (Y) the system will figure out the correlation between the features and the target. You can imagine it like this:
“Hey Computer, look at this weather and sales data I have here. Figure out how the weather is affecting my sales.” (This is called training a model) “And now tell me: rain = 1 and sun = 0, how are my sales?” The system will look up in the previously trained model and return a prediction for the new given features.

Of course, you still need to do some extent of feature engineering. This method is using human domain knowledge to define and set a weight to the features. Maybe the feature “rain” is more important than “sun”. So, let’s multiply the values for “rain” by 2 to give it some more relevance. But you can possibly have hundreds or thousands of features. Which are important, which are not? How about a combination of a few features given circumstance C?  Because of this, feature engineering is difficult and expensive. Some help in this case provides Deep Learning.

As you may think, our predictions will become much more precise. Our own human workload in feature engineering reduces rapidly. But through the nature of Neural Networks, we don’t really know which features in which circumstances are responsible for the decision the computer makes in the end. We only know our data we used for training and the model architecture. Deep Learning can become a Black box very fast.

So, what are the fields of application for this technology? Where does the Data Scientist fit in all of this? Can we see some Code? All of this and more in future blog posts. Stay tuned!