When Natural Intelligence blows Artificial Intelligence out of the water
It is said that “technology is at its very best when it’s invisible”.
If we turn the table on what is ‘technology’ and consider for a second Natural Intelligence (vs. Artificial Intelligence) as the tech in question, we are poised to be blown away on the staggering amount of ‘biotech’ under the hood of our skull.
Tech makes learning and adaptation to a changing world so streamlined and seamless that it is virtually invisible… to the very owner of those machines, us humans.
Only understanding how natural intelligence works, and its wonders, we can really draw a meaningful comparison with AI, and truly realize its limitations.
What’s the thing that continuously happens in our brains without us realizing it?
Let’s look more closely to what happens in our brains on a day-to-day basis.
The ‘computational cores’ of Natural Intelligence are neurons, of which humans own an average of about 100B in a typical brain.
Neurons communicate with each other with spikes (small pulses of low-voltage currents), where a neuron spiking may participate in representing the presence of a feature of an image (e.g., the color ‘red’ in an apple). When neurons tend to be activated together, the connections between them, called synapses, tend to be reinforced. That is why ‘red’ tends to be associated with ‘apple’: their repeated co-occurrence causes synapses between neurons that code for those features and objects to be linked — mechanistic substrates explaining how our statistically-inclined mind work: it helps us find regularity in our otherwise confusing world.
These ‘wiring’ — or learning — events happen all the time in our brains, from the moment they form. Let’s estimate how many times these individual learning events occur in human life.
Assuming an average 100B neurons, 500 trillion synapses, an average 2 billion seconds life for a typical human being, and assuming each neuron in the brain emits a spike every second (this is a low estimate… our neurons are more active than that!), it brings the total number of spikes in the brain that are able to change a synapse (read: cause learning) to a staggering septillion!
That is, the number of times we have a learning episode in our brain is at least equal to, or more than a 1 followed by 24 zeros….. which coincidentally is the estimated number of stars in the universe.
This means that your brain is learning thousands of times by the time you are done reading this post.
Why? Think for a second if this was not the case! Take skill learning as an example, defined as the ability to acquire and improve a perceptual, cognitive, or motor ability as the result of training. Skill learning, a fundamental ability for all animal survival, is evidenced by the fact that humans can improve with practice in essentially every task that has been tested by cognitive and behavioral psychology to date.
While the time scale in which learning occurs varies between minutes to longer time scales, basic neural and synaptic mechanisms outlined above are at the basis of the underlying learning machinery. Our neurons incessantly fire and re-wire synapses in our brains.
Let’s take a mundane example: learning the visual identity of an individual. One may be erroneously thinking that all you need to do is to look at, and learn, a picture of a face and you will be done. But, unsurprisingly, people change over time (see picture).
In the example, if you had met Donald Trump in his early 30s vs. in his days of US Presidency, you may have a moment of pause and dissonance, but your brain would rapidly accommodate and ‘encapsulate’ this new, evolved image of Donald Trump in your personal representation of the US President.
This example of continuous tweaking one’s representation — were these perceptual, but also motor and cognitive — is the rule rather than the exception in our brain and lives. Our homes, workplaces, people we interact with, skillset we are required to use an update, all change over time: nothing is really stable.
While biological evolution has figured out how to design our Natural Intelligence to cope with reality, today’s Artificial Intelligence has definitely not.
Designing and field AI that can exhibit this “continual” (or lifelong, or persistent) learning is an unsolved problem, where AI can be trained to achieve high levels of performance when presented with predetermined datasets but cannot learn continually in settings that are typical of the world inhabited by humans. In our world, the data that comes to us is not neatly prepared in datasets, but comes to us unexpectedly and without warning. Nobody tells us ‘listen Bob, I am going to present you 300 images of tomatoes, learn it’. Tomatoes will come and go with no warning… We will need to learn them all!
Ultimately, machines and more in general AI systems that will need to co-exist in challenging deployment scenarios alongside humans will need to exhibit the same, staggering ability of humans to learn during their lifetime.
Currently, traditionally designed AI cannot do that, limiting the ability of machines not only to work alongside humans, but to challenge us to the highest step of the Intelligence podium.