How do machines learn?

Learning is a process where we acquire new experiences or skills and then apply them in practical life. Let’s compare learning between humans and AI. Young children learn independently. They experiment with their own bodies and play with objects. They learn to walk, talk and recognize objects and people naturally by themselves. 

Artificial intelligence learns similarly. Nowadays, AI is able to learn thanks to machine learning. Most often, the robot learns by trying. It continues to work according to the collected data. The more data it has, the more accurately it is able to correctly deduce the result.

Check Camilla’s Instagram!

Supervised learning

Machine learning with a teacher clearly defines the input and output parameters. The teacher (= scientist) names a certain thing, for example a car, and “shows” how to deal with the given information. The output is then compared with the original expectation. This approach uses rich data sets (often compiled manually) that show different variants of an object so that the machine can subsequently recognize them and perform a given action with them. 

We use this method when we want to predict or explain the data. For example, we can use this method to have artificial intelligence help a company predict the number of new people who will sign up for a service it offers in the next month.

When the algorithm encounters new data, it proceeds according to the set parameters, constantly adjusting the current information. Machine learning trained in this way can only ever do what we teach it. For example, it cannot perform an activity that has not been defined before, i.e. it isn’t able to work independently.

Unsupervised learning

The unsupervised machine learning method, in contrast to supervised learning, only defines the input data. The algorithm comes up with the procedure, solution and result by trial and error. The machine thus looks for similarity and connection between individual objects and assigns them its own marking (tagging). 

Unsupervised learning looks for ways to relate and group a body of data without using a target variable to predict the data. It evaluates the data for properties and uses them to create groups of items that are similar to each other. The unsupervised learning technique works, for example, to help a retailer who wants to group together products with similar characteristics.

In short…

Machine learning can be summarised as following:

Data collection and preparation. The algorithm identifies the sources and creates a structure based on the compiled data. Then the data will be divided into two segments – training and testing models.

Training using training data sets = tuning for maximum processing speed and accuracy.

Testing (Verification) = with the help of i.e. test kit that evaluates the effectiveness of the algorithm.

After testing, we get results (output), such as conclusions or predictions of further user behaviour and machine performance. The entire process of collection, training and validation is constantly repeated as new data arrives. 

Can they learn more naturally?

“Machines that learn like babies give us great insights into how the mind and body work together to acquire knowledge and skills.”

The question of how both humans and machines achieve this with almost no guidance is the centre of developmental psychology and robotics. Their mutual cooperation leads to remarkable findings in both fields.

In prediction-based experiments, roboticist Jun Tani tested how well algorithms would perform on robots learning basic movements. He discovered that machines can learn basic skills such as imitating hand movements and obeying basic commands such as “point” and “hit”. [1]

What is deep learning? 

Deep learning imitates human thinking using various combinations of data inputs. This data works together to accurately recognize, classify and describe different objects. Mathematical models are the core of learning.

Deep learning requires a large amount of data and a lot of computing power, because it sets many parameters to often huge architectures. Therefore, this technique requires a very powerful computer with a graphics processing unit (GPU). This technique is very successful, for example, in the areas of image, text, sound and video classification.

The robot acquired skills that mimics human behaviour with the help of deep learning.

Get to know iCub – the robot that learns independently

The University of Plymouth in England has developed iCub, an android that has been equipped with a neural network that makes it able to learn words. University researchers found that the iCub learned words more easily when the names of the bodies were associated with specific positions. The experimenters repeatedly placed a balloon or a cup on the robot’s right or left side. It could associate the objects with the movements needed to look at them (turning the body or tilting the head). It then matched these activities with the subject names. The robot learned these words better when the object appeared repeatedly in one particular place. Also, the neural network showed the output numbers more accurately when the iCub learned to count on its fingers and not just the names of the numbers. [2]

This is iCub. The world’s smartest robot toddler. Source: Flickr.com

References:

[1], [2] KWON, Diana. Self-learning robots. Scientific American. Los Angeles, USA, 2018, p. 28-31. ISSN 0036-8733. Czech edition.