Scientists and experts often look at the human brain to gain inspiration for building machine learning systems and artificial intelligence. Neural networks, as their name implies, are based on a similar architecture to that of brain neurons, where proteins, fats, and neurotransmitters are exchanged for weights and biases wrapped up in functions (obviously this is an oversimplification, but if I've learned anything from Mr. Feynman, it’s that analogies are powerful tools for understanding). As a result of this biomimicry, advancements in machine learning (ML) and artificial intelligence have exploded in recent decades. With all the focus spent on how to improve ML algorithms, the potential for applying ML strategies to human learning has gone predominantly unnoticed.
Now, what I am about to discuss is, by no means, groundbreaking stuff, but I believe it is interesting nonetheless. It has to do with our ability to learn a thing or two about learning from ML algorithms.
How to Teach a Machine
When we want to teach a machine learning algorithm to do something, say for instance, recognize images of cats, we need to show it thousands (sometimes millions) of labeled pictures of cats. Why do we have to do this? Well let's look at a few examples. (If you're looking to dive deeper into the mechanics of neural networks and machine learning, check out this awesome video created by 3Blue1Brown)
In the first example we'll show our ML algorithm a picture of a brown and white cat laying down. No one would argue that this isn't a picture of a cat, but some may argue that this isn't a very good picture of a cat. Why? Well for starters, if you've never seen a cat before you might be likely to confuse a cat with a snake since neither the cat in this picture nor a snake have visible legs (spoiler, snakes don't have hidden legs either if you were wondering). Right, so it seems like we need to show the ML algorithm a few more cat images.
Here we have a better representation of the average cat: all four legs are visible, whiskers are propped in their feline glory, and gait is nice and sassy. If these were the only two examples we've ever seen of cats, what are some generalizations we can pull? Cats are oblong objects with either zero or four legs, fur (whatever that is), whiskers and sass. Great! Now we can identify any cat, right? Wrong. Although it might be hard to find a dog with feline sass (and I have seen some pretty cat-like dogs before ), it would be fairly easy for us to confuse an image of a dog with our understanding of a cat. Both are four-legged oblong objects with fur and whiskers; what's the difference? Well, that's an interesting question. What is the difference? What is a cat? We can show the ML algorithm a few more images of cats but does it then truly understand what a cat is? (Do I even understand what a cat is?)
Generalization and Abstraction
This issue comes down to generalization and abstraction. Any individual image of a cat is not a cat, rather it's a specific instance of the abstract idea of a cat. In our minds, there exists an abstract notion of a cat which encompasses the true nature and variability of all (or most) cats but is not one individual cat. Why? Because each instance of a cat shows a slightly different angle on that truth. Cats can be big or small, hairy or hairless (although hairless cats shouldn't exist IMO), brown or beige, sassy or wild, you get the idea. Given the potential for variation from cat to cat, not to mention image to image, the only way to make generalizations is to look at enough pictures of cats to see what they all have in common. This step is powerful and important, especially for human learning. We've actually done this quite a lot without noticing.
Thinking back to primary school, many of us will probably remember doing different types of addition problems.
1 + 3 = ?
7 + 9 = ?
2 + 6 = ?
Which one of these is a generalized addition problem? Well, none of them are. They are all specific instances of addition problems. By doing enough addition problems, a student may begin to learn that any number can be placed in either the first or the second position (as opposed to subtraction) and, therefore, could be represented by something that represents any number (a variable). Therefore, a more generalized form would be the variable representation of the addition relationship,
A + B = B + A
The interesting thing to note is we are no longer dealing with the physical quantities or characteristics we saw in the examples but rather their positional relationships to one another. This generalization can lead us to deeper understanding about addition: that addition is commutative (and associative if dealing with more than two numbers). This step in understanding relationships between values is crucial because it allows us to form understanding through generalization, which becomes even more important with complex systems.
If we continue with the cat example, and apply our understanding of generalization, we can extract an even more interesting phenomenon. That is, the true understanding of a cat does not lie in memorizing the specific dimensions and metrics of a cat, but rather, it lies in understanding their relationships with one another.
Examples in Finance
This idea exists in many other fields too. For example, in finance, what classifies an investment opportunity as a good investment varies wildly depending on the circumstances. If we own stock in a company and are given one metric (for example revenue) to determine if it will be a good or bad investment, it would be nearly impossible. The reality is it would depend on (be related to) many other metrics of that particular company. Being a naive and ambitious young trader, one might be falsely led to believe that "good investments" are based on just revenue and cash flows after watching his/her boss execute a few trades that were seemingly based on those metrics. This is partly the reason why "experience" is valued so highly in the workplace—i.e. how well can you recognize the relevant variables in a given situation and how well do you understand their relationship with one another—i.e. how accurately tuned is your biological ML algorithm?
All of this comes down to a few simple actionable steps. When we are unsure of the bias that exists in an instance, the only option is to look at many instances. That is, if we believe the example is not a perfect representation of reality, we need more examples. This is the key to learning like a machine: learn from as many sources as possible, the more varied, the better.
This is the key to learning like a machine: learn from as many sources as possible, the more varied the better.
When learning about current events, don't simply believe the statements of one media company, listen to the statements of the opposing media company too. This will help you discard irrelevant information faster.
Learning about European History? Learn from European historians, American historians, Asian historians and others. Again, the broader the perspectives, the easier it will be to identify the biases and differences, allowing you to take one step closer to understanding the reality.
Learning like a machine is simple but not easy. It requires us to learn from varied and opposing sources, sources that are controversial and/or go against our way of thinking. It can be uncomfortable at times, but when it comes to learning, machines don't care about comfort and neither should you.
When it comes to learning, machines don't care about comfort and neither should you.