Why these 5 reasons convince you that machine learning predictions are lost

Do not be fooled by fantastic predictions

Photo by Michal Matlon in Unsplash

Machine learning has improved our lives.

In general, we no longer have to watch, listen, or even eat.

However, these are only possible recommendations and not as many individual recommendations as you might think.

After thinking about machine learning for a while, I realized that machine learning predictions are like the recommendations you get from a McDonald’s calculator staff – “Do you want that potato with it?”.

Sometimes yes. Sometimes not.

In another note, I can say that our reliance on machine learning predictions is similar to how King David slept with Bath. At the time, this seemed like a good idea, because Bath-sheba was beautiful, but David didn’t really know who she was and the consequences of her actions.

The results of machine learning seem magical, but we don’t really know much about the data or learning algorithms that lead to results, and we certainly don’t know the consequences of blindly following machine learning recommendations.

The ability of the machine learning algorithm is based on the recommendation of its training package. Basically, the algorithm spits out everything that is put into it.

For example, say you want to predict the winner of the next election. If a particular party has won a particular seat in three consecutive rounds and your data is based on previous votes from 12 years ago, the algorithm naturally predicts the same party as the next winner of the course.

However, let’s say that 5 years of social development has changed the demographics of the voting community in this particular seat. How did you predict the winning party?

You really can’t.

A change in demographics does not necessarily mean a shift in loyalty from one side to the other for the region. You can also take a poll and guarantee that if someone wants to vote for him or her correctly during the poll, he or she will not suddenly vote left on election day.

Finally, machine learning can predict the past using past data, but it cannot actually predict the future with past data.

I once heard in a podcast that someone has built a machine learning system to help judges determine the sentences for criminals. However, what the system did was that it recommended the arrest of African Americans at a higher rate than white Americans.

Why is this?

How can a model be biased?

Mathematical models should not take race into account.

And, no, they shouldn’t do it if there isn’t already a bias in the training package.

And, so it was.

The set of historical studies contains a bias of the past, as African Americans were incarcerated at a higher rate than white Americans. Or in other words, the results of a machine learning algorithm are likely to lead to the most common scenario in a data set.

If history is biased, then the machine learning algorithm remains biased because it doesn’t know better.

Personally, I don’t think machine learning predicts anything, but it does automate what is already known.

Perhaps I will put a definition here to explain my point.

A prediction is a prediction – a prediction of the future.

However, what makes machine learning good is that it takes past information and gives you the expected results of the past.

So, really what makes machine learning good is the automation of past thinking. By this I mean that if a data set is already a few seconds old, does it reflect the present or represent the past?

For example, get a house price forecast. What is the real basis of the prophecy? You say interest, land size, location and so on.

However, the reality of this prediction is how one can pay for this house, based on what others have paid for the houses in the past. You can’t really assume that all buyers are wise in making a decision. For example, people do not buy certain houses based on their number, while some buy a house because they believe that it predicts the future cash flow of the house, while others do so because the house is close to the family.

Also, keep in mind that the study of controlled machines is nothing more than the weighting of variables to predict the intended outcome.

Thus, a real-world home prediction is a computer that independently weighs variables and estimates how much a particular variable contributes to the target result, and it automatically performs that assumption based on the weight of the variables each time there is a data change. does. fed to the model.

Furthermore, machine learning models do not actually predict anything when completely new information that is not available in the training package is entered into the model.

I read in this book Against instability by Nasim Taleb. The author writes that it is harder to capture radio waves in complete silence than a little noise in the background. I’m not so sure of the facts, but I think it has something to do with the equipment that makes a difference when the sound is louder.

This example is reminiscent of a computer vision model that named an African man as a gorilla instead of an African man. The machine learning model was suddenly offensive, but behind the scenes the machine learning engineer should have known better.

This tells us that computer vision and machine learning models require not only true positive examples of something, but also false positives, false positives, and true negatives.

Why is this important?

Before you can trust machine learning predictions, you need to make sure that the basic data set reflects the real world. In other words, whether a machine learning engineer just got good results from cherries, the probability of over-matching you with invisible data is ridiculous.

For example, the abandoned Google Flu Trends failed because the car’s learning algorithm eventually linked the flu outbreak to the upcoming winter season. It could not distinguish the false positive events and the real events that led to the outbreak of influenza. “Information Detective” by Tim Harford gives an example that I would describe: “The outbreak of the flu last year in a specific basketball game does not mean that a basketball game next year in the same region and time will also result in an outbreak of the flu.»

Mechanical engineering is not a dynamic prediction machine. It works best in a static environment. By this I mean that if you understand how machine learning works, you will understand that the model must be trained in a set of trainings before it can be predicted.

The situation is that the learning set is static at the same time, and in order to maintain a constantly changing landscape, the model must be constantly developed with a new set of learning.

For large companies, this may not be a big issue. Apparently, they have the computer power to sustainably study machine learning algorithms in new training kits and replace “production” algorithms as training training kits become obsolete. * There may be other ways to learn machine learning online that can solve this problem, but that is beyond the scope of this article.

Instead, let’s say you want to build a system to predict the value of your home, so your results are at the disposal of what the data set was at the time you studied your model. When the number of home sales occurs, it changes your predictions each time, leading to controversy, what is the real value of the home?

You may or may not have noticed, but home cost forecasting programs usually give you a wide range of home value. Typically, it’s a $ 100,000 range. Believe me, you don’t need an algorithm to tell you that. You can easily search the history of homes sold in the area to make such assumptions.

As you can tell, I don’t believe I take the results of machine learning very seriously.

It is really important to know a set of basic training data and sometimes even a machine learning algorithm. * Probably, if you subscribe to the free lunch theorem, the type of machine learning algorithm doesn’t matter either.

I believe that machine learning has a place in society. For example, it’s easier to recommend what to watch on YouTube, even if it’s a long shot. However, in reality the machine learning algorithm instead of actually doing what you want to watch automatically recommends you similar types of videos to watch. I know because I’ve been given recommended videos that I don’t watch very often.

To complete this article, take the results of a machine study with a piece of salt and always ask how the results were obtained.

Leave a Comment