What benefits can we have from Machine Learning?

Hello people. Nowadays there is a lot of hype around AI and Machine Learning and it may be fonded, but let's investigate into it.

Sir Arthur Charles Clarke said:

"Any sufficiently advanced technology is indistinguishable from magic".

Indeed, today many people think about ML as something magical.

But, let's look a bit deeper what it really can do. First of all, we have to remember that ML can answer only 5 different types of questions.

1. Is this A, B or C?
2. Is this strange?
3. How much?
4. How is it organized?
5. What do I have to do after this?

Now let's look into the details and the examples for each of the question above mentioned.

1. Is this A, B or C?

In this case, we'll be using classification algorithms and the most simple form of the classification algorithm consists in telling whether something is A or B. In more advanced algorithms it can choose from more than two options.

Some applications in the real world can be:

• Will this tier hold for 2000 km more? Possible answers: Yes, No.
• What will bring more customers? Choose an answer between, 10 € coupons, 20% discount, 5€ and 10% discount, buy 2 and get 3.
• Is this a cat or a dog? Possible answers: Yes, No.
• Is this a human or a robot? Possible answers: Yes, No.

2. Is this strange?

In this case, the answer will be from an algorithm which is specialized in detecting anomalies,

Some real-world examples can be:

• Are those sheep all mine? is there any intruder? Possible answer: Sheep X is strange. Go check it.
• Are those credit card transactions legit? Is there any weird pattern? Is the card being used in an improbable place where the owner would not go? Possible answer: Transaction X is weird or a fraud. Card blocking is necessary.
• Is the sound of the engine right? Are there any strange noises? Possible answers: The engine is doing a strange sound. It should be checked. An improved version would tell you what are the possible issues also, thus saving precious time to asses the problem.

3. How much?

To answer this question, a regression algorithm will be involved. Those algorithms output a number which predicts something. 

Below we have some examples:

• I have a diamond of 3.4 karats, what is its value? Answer: the average price of the diamond.
• What temperature will be the next Sunday? Answer: The predicted temperature.
• How many ice creams will you sell in June? Answer: The predicted number of sales.

4. How is it organized?

This question requires an elaborated answer and is harder to process because neither you nor the machine knows what are the possible outcomes. In this situation, the grouping algorithms will come in handy.

Indeed, the answer will be a division of the data in specific groups with specific traits and there is no right or wrong answer. 

To better understand this concept let's look at some real examples:

• Let's assume we want to figure out what scooter models have the same type of issues signaled by the customers? The result we will get is a series of groups with the issues and inside it, there will be a series of scooter models which are part of that group. In our case which shares the same issue with other scooter models.
• Next, we can think of a fisherman which wants to know what kind of fish is better to fish in a specific month and where. The answer will be a list of the months having as group members a list of fish and the places where there are the most of them.
• Another example can be the situation in which we want to find out what kind of people go shopping in what areas and when. The answer will be similar to the others, with the difference that in this case, we will get a list of people divided into categories and places they will likely visit in a specific time span.

5. What do I have to do after this?

This is the question which also humans pone themselves and through which they learn. It is the base question which allows an entity to behave in a way based on its experience. The technique used to answer this question is Reinforcement Learning. And the algorithms implementing it, have been inspired by the way our brains answer to reward and punishment.

Indeed, these algorithms learn from the result of past actions. They accumulate data while they live and learn from mistakes. 

As you can deduce, these algorithms are perfect to be used for self-driving cars, autonomous robots, drones and so on.

Some real-life examples can be:

• A self-driving car at a traffic light: it is orange, should I accelerate or should I break? The immediate answer that comes to us in mind, is to break, right? Well, not always, because it depends on many factors, as what is the car's speed, what is the distance from the traffic light, etc.
• Patrolling robot: I still have some blocks to check. Will my battery be enough to engage the defense system if an intruder is detected, or should I first recharge? But how urgent is the situation? How long should I recharge to optimize the time I need to finish the patrol? If you focus on all the little choices and answers this robot has to answer before deciding what action to take, you'll quickly understand that is not as trivial as it may seem at first sight.

And here I end this short introduction about the potential and the limitation of Machine Learning. Stay tuned for a more in-depth analysis featuring each one of the five questions the ML can answer.