A Novel Framework for Creating Self-Learning Artificial Intelligence
Why Biologic Intelligence is the solution we’ve been waiting for
I. Getting Caught Up With Modern Approaches
After being away from the research on AI over the last few years, I took a deep dive over the last week into the current state of affairs.
I downloaded and worked through Tutorials on Theano, Tensorflow and Torch. I read an AMA with Facebook’s/NYU’s Yann LeCun. I read a number of scientific papers from arXiv and a book on a roadmap for machine learning, which you’ve probably come across. I tweeted a bunch of notes and thoughts while I was training my own Deep Learning model with new data (aka my human brain).
And so what was I looking for as a result of all this research about learning? Namely, where the edges of our understanding as a scientific and startup community is. And even with the recent announcements and advancements in Deep Learning, we’re basically still limited to pushing 100s of millions of points of data through a bunch of connected regression lines and then trying to reduce that web of connections into something smaller.
So basically, link one equation up to many others and then just push data through the system. It’s a script, it runs once, and takes a long time. So you need powerful processors like GPUs to do all those calculations in parallel.
Funny how the dots only connect looking backwards. I remember doing this back in the early days of AWS (2007 to be exact) when I had a high frequency algo trading firm and we were using MATLAB and AWS to do stock price predictions over historical data while trying not to over-optimize, then make trades using that prediciton. This is essentially the same thing.
Artificial Intelligence is just an equation.
II. Defining a Neuron
If you listen to the folks talking about machine learning, you hear them all talking about Neurons. And that makes you scared because you remember the nightmare that was Organic Chem from freshman year. And you also see equations and you get scared. I’m here to tell you that everytime you come across something that people try to explain in a scary complicated way, it means they don’t truly understand it.
Don’t let them fool you.
It’s nothing more than a simple equation from Algebra. Remember y = m*x + b? It’s the equation for a line.
Now, take that concept and expand it with the chaos that is the real world. Lots of stuff happening that you’re trying to Organize & Simplify (and this). Just plot a bunch of these (x,y) coordinates and see what happens. You might find that they end up centering around a structure. Like the price of a house or apartment increases the more space you have. 1000 square feet is cheaper than 10,000.
What we’ve just described is a regression line:
You’re now caught up with the most advanced machine learning concepts in the world. You were creating AI in your algebra class in 7th grade and you didn’t even realize it. You’re smarter than you give yourself credit for. You probably just had a shitty teacher.
And so, a neuron is nothing more than a regression line. Said more simply, neuron = equation.
My ‘Neuron’ Is Actually A Deep Learning Architecture
Okay, lets take this concept of a neuron and zoom out. Remember, all of life and mathematics is nothing more than a fractal. It’s the same shape whether you zoom in or zoom out.
And so in the architecture diagram you see above I’ve labeled the dots as “neurons” but what I’m actually referring to is a Deep Learning matrix of interconnected neurons that solves a specific problem. For example, the “Organize & Simplify” piece takes a bunch of existing data and then tries to find common patterns and extract that information, returning the correct approach. It’s an algorithm within an algorithm.
So, we’ve got to take all the research and algorithms already in place and start connecting those together. That’s the basis of this core architecture.
III. Creating the Core Architecture
Our state of the art methods for Deep Learning don’t learn much at all. You just push a bunch of data through a formula and let it work out the details and differences.
So, the next step then, is something that’s self-learning. We are not there yet. The concept, again, is very simple. Instead of this script running once that humans build, you start with something small, turn it on, and let it build itself over time.
What you see in the image below is my approach for a high-level framework. It incorporates gamification (triggers, actions, and rewards), deep learning approaches (connected neurons), and a starting point for setting it off.
It all started with the fundamental concept that it should start with a very simple set of instructions that when turned on, creates an emergent and self-replicating system.
Of course, if you’ve done any research about the universe we live in, and in different concepts outside of just technology, it helps you make the creative leaps necessary. And so, you get to the Rule of Threes that shows up everywhere, including the strongest shape for building solid structures in the real world: triangles.
The image above shows how you start with a very simple triangle, 3 points and 3 lines. That gives you region #1 shown above. Then you add 1 more point outside of that triangle and connect it with 2 lines. That gives you region #2. You keep repeating the process ad infinum and you essentially begin creating squares made of triangles. It’s simultaneously the strongest and simplest structure in building our physical world.
So, then you make the leap that if you can’t do any better than that, and our brains are just interconnected neurons, then why don’t we use the same concept to build out our own digital Connectome.
IV. Creating A Self-Replicating System
So now that we have our architecture in place, we need to create a system that enables data to flow through that system. Remember, the fundamental rule is that this system keeps running and doesn’t run just once.
And so that means we need a force to push or pull data through the system. We need an incentive. We need a reward. And so that means you need to program an innate need and fundamental behavior into this system.
This is why I don’t feel that Terminator or the Matrix or the doomsday folks have it right. If we create a machine in our image, and at the core of that machine we give it the need to learn, then that means its fundamental reason for being is to learn. And that means it will continue to need new forms of data to test hypothesis and create novel connections. In essence, we will have created a machine scientist, fascinated with learning.
Now, have scientists destroyed our world, or has it been the power-hungry capitalists who have taken that research and created destruction from it? I’d argue the latter.
And so, if you look at our architectural diagram, we have a core, central “neuron” (which, remember is actually a deep learning matrix that solves the major question) around which everything else is built.
So you start with the following process:
- Curiosity: Core drive of curiousity that keeps learning at the core of the AI (so it’s less likely to be destructive). Basically, “playing” is fun because you’re learning. That’s the reason for existence.
- Organize & Simplify: all the information coming in, but also its own architecture so it doesn’t grow inefficient as it grows in size. Essentially, scalability in place. Said differently, train it to build its own microservices (another popular software engineering architecture).
- Expectation & Prediction: use the data and learned experience to make predictions about what you’re learning. Or, looking through a different lens, the Scientific Method.
- Choice: this is the fundamental driver of human behavior. Before any action, there is always a choice. This is well understood by the greatest storytellers in the world, especially the main character. What’s interesting from researching choose your own adventure stories is that we should actually choose to follow a character, a human, and let the story unfold, rather than following a cold-decision. This is the art part of the brain.
- Try & Compare: what happens when we make a prediction? We want to see if it comes true. This is the same concept for the deep learning ‘neuron’.
V. Where’s The Magic Moment?
If you’ve spent any time looking to product for software or apps you’re likely familiar with the Magic Moment. It’s the moment of insight and perspective shift where you emotionally “get” the product and the value comes into full perspective. For Instagram it’s when you apply that first filter and make your crap photo look pretty. For Twitter it’s the first time you get a retweet or reply. For Facebook it’s getting 7 friends in 10 days.
So where then is the magic moment for our humanoid little friend? It’s the same concept. When it has a creative leap, a new insight, connects two previously unconnected things, and has a perspective shift of understanding. This is how we learn, and it’s how babies come into the world. Here are some of my notes on the concept from back in 2013 when I was doing deeper research in this area:
As far as I can tell, I have yet to find any scientist or researcher who has written the equation that allows a creative leap to emerge from the underlying mechanical system.
But I think I’m onto something here. If you know something I don’t, please chime in so we can help each other!
This new insight is the last piece that accelerates the learning, it’s the dopamine reward for solving a previously unsolved problem (like getting a software program to work for the first time or nailing that kick-flip you’ve been trying for weeks). This is the new section that gets built at the top of the architecture (the part with the dotted lines). So, as the data keeps moving around this system, new ‘neurons’ are created, and then creates an incentive to keep going, to keep collecting data, and to keep learning.
The question that then needs to be answered is whether that a machine interested in learning is a threat or not.
And so, if it is, then maybe we need to program in the need for rest and entertainment. Find a joy algorithm that lets it watch movies for the enjoyment rather than for the data.
But at the end of the day, isn’t that just a different kind of data? Positive emotion from a fun time?
I’d love to hear from anyone working on this, or has thoughts about this approach. Email me at seanmeverett@gmail.com.
— Sean
