Large parts of our working economy are now fuelled by AI and deep learning. But we haven't cracked general AI - an artificial intelligence that is able to understand and learn any task that a human being can.
This is the event horizon upon which all AI research pivots and has been the subject of countless science fiction stories. It's also one that Google's very own AI company, DeepMind feel like they are on the cusp of touching.
But is General AI really a dream worth chasing and just how close are DeepMind to catching it?
We're likely still years or decades away from general AI. AI systems are still struggling to replicate even the basic cognitive functions of a human child, let alone the general problem-solving capabilities of a fully-developed adult mind.
Rethink Robotics was one of the original leading lights of the AI sector, but it was shut down in 2018 because it simply couldn't settle on a viable business model. Then there's the furore over self-driving cars to unpack, with the infamous self-driving Uber accidents and general uncertainty around the strength of the technology, pushing back the dream of a safe self-driving future quite some way.
While we have managed to make some pretty significant strides, we've still only managed to perfect narrow AI. This is AI that works very differently from the human mind and might be great at performing specific tasks (playing a game or making predictions, for example) but can't process language and visuals as flexibly as the human brain can. This means you have many different AI's focused on individual tasks but none that can "do everything," as it were. It's clever but it's not human-level.
Even advanced concepts such as self-driving cars still fall under the banner of narrow AI. In fact, it's this AI that is at greater risk of replacing human jobs. But narrow AI only works in a limited context and a general AI that can essentially replace a functioning human brain is still decades away from becoming a reality. Or is it?
While general AI certainly has its cheerleaders, others think it's the quest for it that heralds the greatest tools, not necessarily the end goal itself. There is an argument to be made that we don't really need to duplicate humans. Instead, we should be using AI to help us and work with us to achieve together what neither could achieve alone.
In his famous book, Artificial Intelligence: A Modern Approach, Google's AI pioneer Peter Norvig says: "Work in AI has pioneered many ideas that have made their way back to mainstream computer science, including personal computers with windows and mice, automatic storage management and key concepts of symbolic, functional, declarative, and object-oriented programming."
If we hadn't been chasing the dream of general AI then we wouldn't have stumbled upon some of the most important advancements in modern human history. The research itself teaches us so much about the laws of intelligence and how to apply that to everything, from menial tasks to predicting medical anomalies and even creating art!
Even the great search for the autonomous vehicle has led to technologies such as auto-parking, lane assist and drowsiness detection - advancements that have saved countless lives.
DeepMind is a British AI company at the forefront of what is known as "deep reinforcement learning." These are the kinds of AI systems that might be able to win a game of chess but are unable to apply what they've learnt in a real-world environment due to sheer unpredictability. Recently, however, DeepMind claims to have taken the first steps in training an AI that can win multiple games without human interference.
It's an "open-ended learning" initiative that places deep reinforcement learning agents inside a 3D engine. The idea is that the AI will learn similarly to animals and humans; by interacting with its environment. The 3D engine is being dubbed "XLand," and can generate environments composed of static and moveable objects.
XLand is capable of generating vast arrays of unique environments and challenges to train AI agents in a manner far beyond more traditional engines such as the ever-popular Unreal engine. The engine allows DeepMind to create literal billions of tasks across varied games and environments, with new tasks being generated continually for the AI to overcome.
According to DeepMind researchers, their agents have "been able to participate in every procedurally generated evaluation task except for a handful that were impossible even for a human," and yet they are only just getting started.
Already, researchers have recorded evidence of "high-level" activity in their AI subjects - "heuristic behaviours" like teamwork and multi-step planning that would have been thought impossible only years ago. This could have major implications on AI problem-solving, which means safe self-driving cars (that old chestnut again) might actually be within our reach.
As encouraging as these results might be, Google is hesitant to jump to conclusions as researchers admit these behaviours may simply be accidental. But they do feel their AI agents have been able to demonstrate self-learned skills and that, while it might take another few decades or so at least, reinforcement learning will eventually lead to artificial general intelligence.
DeepMind's research proves that if you give an AI access to a complex enough environment and give it the faculties to learn by doing, then AI might eventually evolve in much the same way human beings have over the last few million years. But the real world is much more complicated than the virtual one.
That chasm might be closing, but we'd be amazed if it was cleared within our lifetimes. But then again, stranger things have happened.
If you've been inspired to discover more about AI and machine learning, then get in touch with Novatech today and enquire about our deep learning workstations.
Posted in Training & Simulation
Published on 18 Nov 2021
Last updated on 18 Nov 2021
25 Nov 2021
Many businesses felt their IT resources were stretched before the pandemic. Now, hybrid working is causing a lot of headaches for IT teams. Join us as we outline a couple of the major frustrations being faced, and how they could potentially be overcome.