To understand how machines learn I would like to draw an analogy with how children learn. Or rather, how newborn babies learn. When children are born they have very primitive innate intelligence like crying to attract others when they are hungry or in pain or want attention.
Of course, doctors and scientists would like us to believe that they have much more intelligence, and I completely agree, but here I want to talk about the broad things.
As the baby grows they learn to identify people and objects around them. They don’t even know what is good for them and what is harmful. You will often find adults catching hold of children running straight into a swing, ride or even a wall just for fun. Without realizing that it could cause harm and they can get hurt.
But then, after some months have passed the same child knows how to keep themselves safe around objects or situations that can harm them.
How has this change come about?
Let’s continue this interesting example of a child learning whether ramming their head into the wall is good for them or not.
They can learn in any of these two ways:
- See someone getting hurt when they hit the wall.
- Their parents or elder sibling or others around them telling that they should not ram into the wall because it will result in a gash or at least a bump, which will hurt.
But children being children, they would want to explore and experiment on their own.
Have you seen a child hitting their head against the wall very softly, then a bit harder and then a bit more till they know the threshold above which the pain would be unbearable? Yes, they do!
You don’t believe me?
I don’t fault you but I speak from experience. My elder daughter was always very careful about her own safety and so telling her that something could hurt her was sufficient. Not so my younger one. She has to learn everything herself, test everything out herself. So I can’t even remember the number of times she hit her head against a wall or a door, scraped her fingers against the scissors, jumped from the bed really hard or tried to skid on the mud to see what happens.
But I digress.
Besides the stories, what I want you to remember is this. Children learn by
- Listening
- Observing
- Asking questions
- Exploring
- Experimenting
Note that the first three methods of listening, observing and asking questions needs input from their environment and people around them. And that is the way machines start learning too.
Eight types of intelligence
Before we proceed further, I would like to point out that humans possess different types of intelligence. To understand how our brain works, psychologists have put forth many theories for classifying intelligence; Howard Gardner’s theory of multiple intelligence is the most acceptable one, and it speaks of eight types of intelligences.
Image courtsy: https://www.cnbc.com/2021/03/10/harvard-psychologist-types-of-intelligence-where-do-you-score-highest-in.html
How scientists teach machines
Machines already have logical-mathematical skills through the various software that computer programmers wrote for them. This excessive obsession with equating human intelligence with logical-mathematical skills in fact led to the failure for AI researchers initially.
Scientists and researchers turned their attention to visual intelligence. Recognizing objects.
When we teach children to identify a shape or colour, we show it to them repeatedly. Similarly, we feed millions of labeled images to machines so that they learn to identify patterns that define objects, faces, plants, even medical conditions.
Programs that train the machines on these input images to achieve the desired outcomes are called machine learning models. A major task of machine learning models is to teach the AI machines to identify patterns. This is done by showing, or providing as input, millions of labeled images. Once the machines has seen enough images to identify defining patterns, it can identify an unlabeled object with high accuracy.
Remember, machines do not start with 100% accuracy. They improve over time as they get more and more input data.
AI machines are helping doctors detect diseases like cancer early on and save lives. This is because they don’t get tired of repeating the same set of processes endlessly, and hence don’t make errors. Also, they can access and find patterns in millions of images in seconds.
Final Thoughts
You might be wondering why do we need artificial intelligence at all, if it’s nothing but a replica of human intelligence. Why invest time and resources?
Artificial Intelligence is required so that machines can do the time-consuming work that assists experts in scaling their capabilities rather than doing repetitive tasks.
This post first appeared on Jul 5, 2021
Trackbacks/Pingbacks