Why are children experts in Machine Learning?

While studying various Machine Learning techniques among Supervised, Unsupervised and Semi-Supervised methods, it became clear to me how babies and kids seamlessly deploy these concepts on a daily basis. Humans are completely unaware of how the world works when they are born, pretty much like a Machine Learning model that has not been trained yet. I gathered some of these analogies below and became even more amused with how the nature is incredible and how we try to mimic it in Engineering and Technology in general.

Reinforcement Learning

The most important thing to learn as a child: how to not get grounded by your parents and need to stay at home while your friends are playing outside! It is all about behaving in the way that your parents expect you to, in order to get a certain reward, which aims to encourage that action from now on.

Takeaway: In the Reinforcement Learning context, there is an agent (the kid) that is in an open environment (life with the parents) and has no predefined data or instructions on what is necessary to maximize his/her reward (time outside playing with friends, candies, screen time or any other prize).

Contrastive Learning

If a kid or a baby is in contact with 3 pets, let’s say 2 dogs and a cat, even though nobody necessarily told him/her that they are “dogs” and “cats”, the child will still realize that the two dogs look similar compared to the cat.

Takeaway: a kid’s brain can learn the high-level features of the objects, exactly like Contrastive Learning does by comparing pairs of unlabeled data to learn attributes that are common or different between the data classes, before even performing a task like classification.

Clustering

Kids always like being parts of groups, those being of similar interests for sports, board games, videogames or any other topic. They are also very quick to identify at school who is part of which group, without necessarily having to clearly label each of them.

Takeaway: unlabeled data is available (children with different interests), because there is no ground truth present (proper denomination such as “group of sports” or “group of board games”) and still they are able to identify intrinsic patterns in the samples (the other children) and therefore make friends.

Generative Adversarial Network (GAN)

Suppose that a child is trying to fool you pretending that he/she is asleep. When this is done for the first time, of course you will clearly notice that it is a trick and that he/she is actually awake. As the kid starts to realize how you find out that he/she is pretending, the fake state will start to be more sophisticated and harder to be identified by you.

Takeaway: the principle is the same of a GAN, which makes use of a generator (the kid, trying to fake as realistically as possible) and a discriminator (you, trying to distinguish between whether the kid is faking the sleep or not). Based on the feedback of the discriminator, the generator will refine its fake sample generation.

Variational Autoencoder (VAE)

Ask a child to draw a dog and although its representation will not be perfect, its main features will surely be present in the drawing.

Takeaway: drawing a dog is like a VAE model, which makes use of an encoder to extract high/level features of the image (in this case ears, nose, mouth, 4 paws), represent it with a probability distribution and finally create a new sample (in this case, the drawing) with the decoder (creation of a new image based on the learned high-level features of existing samples).

Semantic Segmentation

Every kid playing Where's Wally will categorize objects in an image into a certain class or label, mapping it entirely. In the image below, for example, it is possible to identify instances of boats, beach umbrellas, sunbathers, swimmers...

Takeaway: this task is inherently Semantic Segmentation, composed of Classification (classifying an object in the image, where each pixel is assigned to a specific class or object), Localization (finding the object and drawing a bounding box around it, in this case Wally is of interest) and Segmentation (grouping the pixels in a localized image by creating a segmentation mask).

The bottomline is that:

• Children are all the time getting accurate representations of the world based on all kinds of data that they are exposed to;
• Even small children can make new inferences that go beyond their experience, showing that they also understand about Bayes' theorem;
• Recent studies demonstrate that even preschoolers learn through imitation and by listening to what others say.

In other words, our high-tech trends involving lots of Mathematics and high-performance computers are carried out effortlessly and efficiently by babies and children at a very early age!