Sci-fi has painted many horror pictures of the future of machines taking over the world. These grim visions of all-powerful machines are some of the many misconceptions people have about AI and Machine learning.
In this article, our Machine Learning Specialist Juuso Lassila answers the most frequently asked questions about Machine learning and tells us why machines are not quite ready to take over the world.
1. What is Machine learning?
Machine learning means that a machine can learn without being programmed. Machine learning and AI are often mistakenly used as synonyms, but they are not the same. Machine learning is actually a subset of AI.
“AI is quite a poorly defined term. It can kind of mean anything,” Lassila explains.
The big difference between Machine learning and AI is that AI doesn’t always need data. For example, some say that finding an optimal route on a map is AI. This process is not data-based but an algorithm.
Machine learning, on the other hand, is AI that uses data to learn. The benefit of Machine learning is that the programmer doesn’t need to know how the model solves the problem. This means the machine can use techniques that the programmer was not aware of. Today, Machine learning plays an important role in many industries.
2. How is Machine Learning used?
Machine learning tries to understand data and the structures in it. To put this simply, let’s take something that the internet loves: pictures of cats and dogs! So, if you have hundreds of images of your favourite furry friends in a random order, you can use Machine learning to differentiate between these images. The machine can organise the data so that you end up with only pictures of cats, for example.
“Machine learning will predict values from data that you feed to it and make sense of that data,” Lassila says.
A common way to use Machine learning is for the classification of data, like in our example with the pictures of cats and dogs above. However, Machine learning is also used in AI chatbots and to make different types of predictions. For example, Machine learning can be used to estimate the value of a house based on the size, location, the year it was built among other information.
3. How do machines learn?
Machine learning is usually trying to minimise errors of a model. First, it starts with random answers. Then humans give it the correct answers, Lassila explains.
There is a way to measure how wrong the machine was with the answers it gave. In some Machine learning models, differentiating is used to minimise the errors.
With differentiating, you can calculate what direction the parameters should be changed in order to decrease the error function's value.
“Then you know how to change your parameters so that the function’s value decreases. You then take a small step towards a smaller error by changing the parameters. That way you minimise the errors,” Lassila says.
4. Can a machine be error-free?
In theory, it is possible that there are no errors but in the real world, it is not really happening. Even in the best cases, the machine can only get maybe 99.9% of the data correct.
“For classification, we can measure the accuracy,” Lassila says. Usually, 90-95% accuracy is considered quite good. When the error margins are low enough the machine is ready to go into production. How low the error margin has to be before the product can go live depends on the task the machine is supposed to do.
When talking about AI chatbots, it is difficult to say when the bot will be more helpful than annoying since AI chatbots can have as many example questions as possible. The machine always has to pick the best one of the available answers. How correct the machine is depends on how many example questions there are.
“If you have one question, the machine is always correct. If you have two, it is hopefully quite often correct, but if you have 100 questions then the machine is quite likely to pick the wrong answer,” Lassila explains.
5. What is Deep learning?
Deep learning is a method of Machine learning that uses layers in the models. One layer produces an output and that output is transferred to the second layer and so on. Deep learning is usually done in neural networks.
“Neural networks are quite handy because they adapt to different input and output formats well,” Lassila says.
Neural networks are very versatile, and that is why they have many use cases. AI chatbots usually use neural networks, as well.
“We create those sentence embeddings from messages which are just vector points in space. Then we measure their similarities. It uses a neural network. It has a lot of layers and it gives output at the end,” Lassila explains.
There are also many machine learning methods that are not deep. Such as support vector machines or decision trees.
6. What are the bottlenecks of Machine learning?
Data is one of those things that causes limitations to Machine learning. If you don't have data, Machine learning won’t work. You also need computational resources. Often the models cause limitations as well.
“For basic tasks, like classification, the models are good. And in those cases, you are just limited by data and computational resources. But in some domains, the models are not so good yet, “ Lassila says.
The more data the better, as long as you have the resources to compute that data. But, the data has to be of good quality. Also, the models have to be big so that the model is able to handle all the complex inputs it might get. You need a lot of processing power for the machine to handle that data.
“You can have access to a lot of data. That is not the problem really.” Lassila says. “The issue is the models and the time it takes to train them. A lot of memory is needed to store these models.”
Another pitfall of Machine learning is the currently available model architecture for some tasks. There is no good model architecture for sentence embeddings at the moment.
“It is still ongoing research to find a really good model,” Lassila says.
7. How do you teach a machine?
First off, you need a lot of data, and the data that you use to train the models need human labels.
“Big companies that use Machine learning have armies of people who just label the data,” Lassila explains.
Even if the bot can start labelling the data with time, you don't want to use the answers that the model gives you as training data. The models make errors more than humans. If you use the machine labelling as training data, you teach those errors again to the models.
“That will make the errors grow bigger and bigger. “
It is also possible to use active learning where the models can tell what kind of data points would be useful to label. This can make learning more efficient. “The machine can guide the labelling process. Still, humans should do the labelling so that the errors don't propagate. “
Once humans have trained a model with the data, then the machine should be able to reproduce that labelling for the unseen inputs. But if you need new answers then the data needs to be labelled again by humans.
8. What is the future of Machine learning?
Machine learning is developing really fast at the moment. A few years ago another deep learning model, the transformer model, was introduced. The discovery of the transformer model has had a huge impact on Machine learning.
“We have been rebuilding the whole research field around that discovery. It has really worked well,” Lassila explains.
The transformer model is also used in AI chatbots with Natural Language Processing. “But at some point, we will reach the end of the road on how well we can use that architecture,” Lassila says. “Hopefully we will find something new that gives even better results.”
9. Will machines take over the world?
People expect too much or not enough from AI and Machine learning.
“There is a lack of understanding of the capabilities of AI, “ Lassila says.
So it seems that machines will not take over the world quite yet. Bots have a few hurdles to overcome before that happens.
“The models and the architecture are not there yet. We do have the data and the computational resources to make a good AI but we still need those good architectures for robots to take over.”
Even though AI is not ready to take over the world yet, it will take over some jobs.
“Some jobs might be replaced by AI at some points. These would be repetitive jobs that don’t require much thinking,” Lassila says.
There are two ways to think about how AI will change the job market. One is that AI replaces jobs, but it could also enhance jobs. It can work alongside humans. A good example of this are chatbots, which can connect to a live agent when they are unable to assist the website visitor.
“AI can improve efficiency for many jobs.”
However, you still need human resources to combat the inaccuracies. AI doesn’t think like a human. It uses different methods. And sometimes it doesn’t make sense to us why it shows something.
“One of the biggest issues with AI is that the machine does not explain why it created a certain output. It might make sense for the model, but not for humans.“
If you want to find out more about Machine learning and AI, check out our AI chatbot guide.