I talk a lot about artificial intelligence, it's potential, it's promise and it's danger these days. Most individuals I speak to agree that AI will allow for most manual tasks to be automated very soon. They also agree that AI will be able to be infinitely more efficient at doing tasks like transportation, cleaning and construction. With some explanation, they concede that AI will be highly efficient and better than humans at doing tasks requiring some creativity or intuition like guessing, analysis and speculation, because those can be trained using the large amount of data available through the Internet.
Then people have a very hard time even considering AI would be able to be artists. As they say: this is the realm of humans. Only humans can be creative, they say. Otherwise, what would we be doing in a number of years? How would we be useful to society if the last type of activity where we excel at can be done better by artificial intelligence?
Of course, this is a reaction of fear.
The truth is, AI laboratories are already experimenting with deep learning AI and have achieved interesting early results: AI can be trained to be creative and make music.
In the following video I discuss a bit more about the social and philosophical implications of this advancement:
What creative AI needs to get started is inspiration. The sort of inspiration they need is similar to what human artists need really: life experience. In this case, since we're not interested in spending years teaching the AI about life in general, the scientists taught their creative AI about the type of art. In essence, they gave them a huge amount of music to digest, from different styles and then through code and algorithms, gave them the ability to make completely original music.
What these AI are capable of is basically the same as what human artists are able to do: get inspired by what they "know" and create something new. They use algorithms called deep learning algorithm where the AI has access to large amounts of appropriate data for the task and then set to "learn" how to play with the data to get target results. If the training period is successful, the AI is able to get the results we were interested in with the data on their own and can continue to learn over time. Neat huh?
Here are a couple recent examples for you guys to check out.
First is the example of the Marimba-playing robot Shimon from the Georgia Institute of Technology. It is fed four measures of music and then it goes on its own to create the rest of the music on it's own, inspired by both the first measures, which provide a specific desired style, and it's huge databanks of recorded music. See it in action here:
Second is the example of Amper AI, which is a commercial venture that allows artists to create music by providing the AI with a style and tempo, and it'll create music on its own. This is a beta product that has already been used by artist Taryn Southern to make new songs. After the AI has produced the musical score, the person using Amper AI can adjust and layer on lyrics and other pieces of pop music to taste. Here is more information about the effort.
Music made by Amper AI, for example when asked to make music styled after the Beatles (lyrics were provided by a human being):
Here is Taryn's song made using Amper IA, after made into a music video (lyrics and video by Taryn. Instrumental elements are all Amper AI):
With great commercial ventures out and fantastic research on creative deep learning AI, we're sure to see more and more people using it and of course for the technology to become way more refined and advanced in a very short amount of time.
Stay tuned and for those interested, you can start playing with some of these projects yourself!
Ok. Once again, I'm getting ahead of myself here with a title. We're not exactly on the verge of creating an artificial brain, and by that, I mean a brain that is completely designed by human beings.
But thanks to recent research, it is at least no longer an unimaginable thing. And who do we have to thank for this? Artificial intelligence of course!
As always, I talk more about the impact and implications of this in my video (see below), but for the purpose of the blog, let's look at how we'll be able to achieve this feat using current technological advances.
The first key to making an artificial brain, is to understand better how the brain actually works, and that is still a mystery, really. The average human brain has 100 billion neurons and each neuron has roughly 1000 connections to other neurons called synapses. That means 100 trillion synapses firing in some mysterious pattern that, according to expert scientists should explain how we even have a consciousness, on top of our ability to remember, act, have emotions and all that.
Now, neurons are cells and synapses are very small structures, not to mention the thin axons that worm around like tentacles to reach out towards the next neuron of interest.
In the image above, we can see one neuron (purple) reaching out with an axon towards another neuron (green) and connecting to it using several synapses.
Now imagine taking pictures of tiny areas of the brain filled with these in tightly packed formation and mapping the synaptic connections in order to ultimately map how everything is connected together, to create the human neural map so others can study how exactly things fire and work.
This is needed because we can't really understand how the brain works without that info, just like we can't really reverse engineer a computer without understanding how the microchips and transistors are all connected.
The current problem is we are able to take pictures of tiny areas of the brain quite effectively (thank goodness!) but we need hordes of unfortunate science interns to actually manually count the synapses and map them out. Right now, using the best electron microscope available for this, which does the job 50 times faster than the usual electron microscope. Still, because the interns have to map things out manually so far, a mouse brain would still require a few years to map. Since the human brain has 25,000 more neurons and connections than a mouse's, using this method, it would take us, well several thousand years to fully map a single human brain.
So using these techniques and others, mapping of the human brain is in fact within reach likely within the next 20 years by my personal estimate.
Now there are other elements needed in making a human brain, like memory and the actual processing of information as well. The human brain can process multiple pieces of information simultaneously using the same neurons and works like an analog system, not digital. In other words synapses work by degrees of positive, like a gradient, and each neuron connects with other neurons with more than one synapse usually, allowing for multiple streams of information to go through between these brain cells. In comparison, our current computer technology, can only process one piece of information at a time and only in absolutes (1 or 0) through the transistors and using about 100,000 more energy than the biological synapses. Very inefficient.
Well, we're in luck, some laboratories are working on artificial synapses that are analog just like our organic synapses and just like our synapses can learn and adapt to how information passes through, thus consuming much less energy than the current transistors. These new experimental artificial synapses, are only 10,000 less energy efficient than their biological counterparts.
So instead of relaying only a 1 or a 0 as piece of information, these synapses pass any value in between depending on the intensity of the impulse provided and the more information passes through it, the stronger the signal, because the synapse becomes more energy efficient. This gives it memory of sorts and can create behavior changes in the output, just like our brains develop habits when we keep doing the same thing or thinking the same way repetitively.
So, these synapses are the building blocks for a learning artificial brain.
Add to this research our ability to conceive much more efficient ways to keep information and we're getting much closer to unlocking and replicating the wonders of our own abilities to think and even our own awareness of the universe.
A good example of this is DNA memory, a technology in development that would allow us to store millions of times more information in the same space as our current hard drive technology and even replicate itself to tackle multiple memory tasks at once.
Another good example, is atomic memory, another field of research that would allow us to hold several states instead of only the 1 and 0, within a single atom.
Either of these researched techniques will become cheaper and more easily understood over the years and would allow us to record all the information flowing through the Internet today on our own personal computers if we really wanted to. Amazing stuff.
We're on the cusp of becoming the creators of many machines that are just as capable as we are to do many things, as I mentioned in many blogs and videos. Be ready and look for opportunities to dive into those fields. They will only grow in importance over the years as more and more research goes into applications for use by the general population.
The elephant in the room for me is not whether Artificial Intelligent machines or robots will be able to gain sentience, but when.
Since the first software we dared to call Artificially Intelligent, we've been talking about simulating human behavior and patterns. In other words, we were able to copy human behavior using lines of code, algorithms, cameras and math.
It was cute, it became practical and now these bits of code and even machines are integrated into our society either in very visible ways (computer opponents in games) or very subtle ways (predictive software for the stock market or Google's Search engine).
Fast forward to today's research and the tech industry's surge to develop more and more intelligent applications of our artificial intelligent expertise. We are developing and commercializing artificial intelligence able to:
Understand human patterns of behavior in order to sympathize or compete against humans.
Those sound like the patterns of intelligence of a growing and learning child to me. Doesn't it?
In the video here, I talk a little bit more about this idea if you wish to go deeper:
In any case, seems to me that we are passed the theoretical science fiction idea of possibly machines eventually becoming our friends, partners and colleagues. Once we are able to properly roll out artificial intelligence that is capable of interacting with humans on the emotional intelligence level, we can make some that people can get attached to.
That, my friends, is an aware, sensitive, high performance work colleague.... or the AI colleague 5 years from now.
When are we going to start thinking of AI more along the lines of Synthetic Humans (or Synthetic life) versus the current ideology of seeing them as Simulated Humans?
If it looks like a duck, quack likes a duck, and is a better duck than an actual duck....
