Evidence is finally mounting that imagination is a key element of decision making and intelligence

Published: February 26, 2020

In a recent paper published in Cell, a team of researchers led by the neuroscientist Loren Frank of the University of California discovered something truly surprising, which seems to strongly support my speculations about creativity, imagination and intelligence. Read on, and... Hooray!!!

But first thing first, here is a wonderful article about the paper: In Brain Waves, Scientists See Neurons Juggle Possible Futures. And here is the paper itself.

Why this makes me happy

Because this reasearch demonstrates why AI needs creativity and immagination in order to become AGI.

I've always speculated that we run many simulations of reality in our brain in order to decide how to act. Some parts of the brain generate the simulations, others evaluate them and makes decisions. Now there is a paper that discovered this is just what happens in a mouse brain. Hooray!!! So this is the first paper about a scientific research that confirms the hypotheses that I have been supporting for some time.

Many believe creativity and imagination are things unrelated to intelligence. This is false. An intelligent agent first need to be creative and to have imagination abilities, only later it can become intelligent.

Why is this? Well, my speculation is that it's because there isn't a universal formula of intelligence, there are not optimal solutions to every kind of possible problems which intelligent agents can employ. Most of the time reality is impossible to model, the only way to achieve success in real life situations is to immagine and simulate in our brain as many outcomes as possibile of our potential actions, and later decide which of those is what we prefer so that we can decide to execute exaclty the actions that broungth to that specific simulated outcome and not others.

Yes, of course this is not the whole story. Many other studies demonstrate that a reward system is also in place in intelligent brains and most probably this is also adopted to establish which of the simulations our brain runs are more accurate and reliable and which are less. But the central point is that not only our perception systems are all based on the active and creative generation of internal representations of the external reality, sort of unconscious inference of what we decide the world is, but also our decision-making process is indeed implemented on the basis of creatively fantasising about what reality would become if we acted this or that way.

No creativity = no intelligence

All these processes need an inherently built-in creativity engine. You cannot imagine nothing if you can't be creative, none the less possible future outcomes, let alone the actions that could cause them. We should stop believing we can reach Artificial General Intelligence without first implementing agents that are creative and have the ability to imagine and fantasize. The sooner we stop believing this absurdity the sooner big progres will be achieved for AGI. But where to start?

Why GANs mark the way

GANs are the first ever artifacts that implement the logic "let's imagine something and then let's evaluate its quality". This is exactly what GANs make. One part of a GAN creates something, many many variants of something indeed, another part of the same architecture evaluates the quality of what the first part has generated. That's why they are called Generative Adversarial Network. Generative is the part that creates stuff, Adversarial is the part that evaluates the generated stuff, Network is the whole structure. While still very basilar and embrional this structure is truly revolutionary and absolutely the right starting point to investigate and research other more complex structures, but still based on the concept of simulating something and evaluating those simulations. These processes are what are going on in intelligent brains, we should invest in this approach to reach AGI.

As I see it the next big step to be made is to invent a network in which the generative part creates more variants in one shot, so that the adversarial side could select one winner, then Integrate this whole network with a sort of built-in reward system and et voilà I think 80% of AGi would be on its way… Join me and let's do it!