Thousands of generations to learn anything | Evolution

AI, in the form of neural networks today takes hundreds of thousands of iterations to learn from scratch. Given a set of defined goals and a basic know-how of the limited set of controls it has, it repetitively performs a sequence of operations to understand what helps it reach its goal and what doesn't. At times it takes the same path with a slight detour to see if that made any difference.
Take for example the following brilliant demonstration of a neural network learning to play Mario. 
Easter egg - MarI/O is open source.
The entire process of the neural network learning to play the game is based on biological evolution. During its initial generations with almost no idea of what to do, MarI/O simply stood there, or kept jumping in place. It then learnt that pressing the right button helped it reach closer to its goal - to reach as far right as possible. So over the next hundreds of generations, it just kept pressing right, and then jumping in place. Later, it realized pressing two buttons at a time can get it even further right, so it started that repeatedly. Slowly, over thousands of generations, playing the same level over and over again, it started to understand what helps it acheive it's goal and what doesn't. But it took the AI thousands of generations to do that.

Now, humans function in exactly the same way. That statement is an irony - the above simulation is a mere replication of the way we learn over generations. Anyway, we're precisely the same. From inventing the wheel to learning about fire, and later making a combustion engine that powers those wheels, we've taken thousands of generations. Science in the same way takes years and years to evolve. Given what we know so far, we make changes, bits and pieces here and there.

What if we didn't take so much time to learn? What if we learnt at a 100% efficiency form our previous generations? Take pollution from our cars as an example. From previous generations, we know it's troubling our atmosphere. We should stop using them immediately before it gets any worse. But no - we can't. We're programmed to learn slowly - we make small changes first. We make refined petrol that burns better and leaves less soot behind. We then create more efficient engines. We keep making minor modifications to what we've developed in the past, until one fine day we learn about electric powered cars with zero emisions. Analogous to the way MarI/O first learnt about pressing two buttons at a time. We then head towards electric cars slowly, developing cars that emit less polution - slowly, over several generations we make cars with zero emissions. We get closer to our ultimate goal of keeping the human race alive. Why can't we just learn faster?

-

And yes, I'm saying this whole species is one huge neural network.

Comments

Post a Comment

Popular posts from this blog

[Breaking News] AI takes over universe to calculate Pi

If in the next couple centuries we are able to build an AI that surpasses human intelligence, will it exterminate humankind? It simply might as it may foresee humankind turning against it and pulling its plug of life. As a race, we have not been subjected to another species more intelligent than us, and when the day comes, we might find it very difficult to control the motivation of a race smarter than ours. A naive proposal suggests making an AI with seemingly benign goals, one as simple as calculating the value of Pi. But here's how it can backfire. "One popular scenario images a corporation designing the first general purpose artificial super-intelligence and giving it the simple task of calculating the value of Pi. Before anyone realises what is happening, the AI takes over the planet, eliminates the human race, launches a campaign to conquest to the ends of the galaxy, and transforms the entire known universe into a giant supercomputer that for billions upon billions
Continue reading »

Firebase Auth | The Debug vs Release Signature Problem

tl;dr 1. Update Firebase console with SHA1 of your release keystore. 2. Re-download google-config.json file. 3. Place your release keystore in the same folder as your gradle file and update gradle with the code below. 4. Tadaa! Authenticating a user can be frustratingly tricky at times. I use Firebase all the time to do all server related tasks - I used to create my own Python Flask server prior to this. You literally have to write no server side code, and it's all for free. All steps at the Firebase Auth documentation are straightforward, except for the part that requires you to add the google-config.json with correctly configured SHA1 key. I got stuck at this one point, and I'm certain most developers will at the same place. It took me hours to do this the right way. Here's what you do : Firebase Auth requires SHA1 fingerprint of your release keystore. Run the keytool command given in the docs. keytool -exportcert -list -v -alias alias_name -keystore path
Continue reading »

Cold Showers

Image
"I feel tired" I said. "Then stop feeling", he said.  This is one of my fondest memories of my first job. My mentor just casually slipped it in, but it's a constant reminder that I have a choice . - Cold showers. The Internet says it's healthy for your skin. Huberman adds a bunch of biology and says it gives you a 4 hour dopamine effect. But more than anything, I say it's your mind's win over your body. It's been half a year now, but every day I step in to the shower, my body doesn't want the cold water. It sends signals to my brain saying this too much, this is too cold. I start with my feet, go up to my hips. And then my body begs me not to pour any over my chest. My innermost fears of visualizing a shark in deep waters surface the moment I close my eyes. Every single day. And yet when I pour the cold water all over me, I feel great. I realize I can just   decide  the water isn't cold. And then it isn't. I don't shiver, I truly
Continue reading »