The Next Level of Computing?
Here’s a bicycle that was showcased in July this year. It’s self-driving, can navigate around obstacles, and can follow a person and voice commands.
This humble bike probably isn’t so appealing since most of us have witnessed AI such as Siri and robot Sophia. However, its internal structure could hold the key to our future and unlock the next level of computing, one that shouldn’t be confused with Artificial Intelligence (AI) – that’s Artificial General Intelligence (AGI).
AI vs. AGI
So far, all AI solves specific problems. The self-driving bicycle moves from point A to B without crashing and Siri understands and processes verbal instructions given by the user.
But when thrown into uncharted territory, these AI algorithms are utterly useless. For example, most voice assistants aren’t exactly the most helpful when it comes to physical and mental health. Try this out now – tell Siri or another voice assistant “I’m sad” and see if it makes you feel better…
Meanwhile, AGI works like a human brain, being able to generalize algorithms and detect nuances. So far, we haven’t achieved this yet. Even Sophia (pictured below), who can perform face tracking, emotion recognition, physical movements, and is a ‘citizen’ of Saudi Arabia, is not considered to be AGI, even by her co-creator. Does this mean we can also grant our self-driving bicycle citizenship too?
That’s up to you to decide. The real question is, are we even capable of creating AGI? It turns out we have a bottleneck and it has to do with the difference between the architecture of human brains and computers.
The (very old) computer architecture
This is the standard (Von Neumann) architecture of computing and has been for about 75 years.
‘Von Neumann Computer Architecture’ by me (Canis Nugroho)
Every computer needs input and output devices. You have mechanisms such as keyboards and mice (computer mice, not live ones) to feed your computer instructions and screens for the computer to return your commands.
The meat of the work, however, comes from the memory unit and the control processing unit (CPU). The memory unit stores the data and instructions in the computer, whilst the CPU processes and executes the instructions. Within the CPU, you have an arithmetic logic unit which executes instructions, and a control unit, which is essentially the boss of the CPU.
The problem here is that the CPU and memory are stored separately. This means that information is constantly being transferred between the two, as shown in the diagram above.
But this is not how the human brain works! And this is why we’re still faster than AI in so many tasks such as facial recognition. We don’t have all our memory on one side of our brains and do all our processing on the other!
The other issue is that this architecture takes heaps of energy because of this constant transfer of information between two parts of the computer. Studies from Computerphile’s neuromorphic computing video suggest that if we continue with our current structure of computing, we would need more energy than what’s currently available in the entire world.
Clearly, it’s time we unite that CPU and memory together.
A brain-like structure
And this is where the hero of our self-driving bicycle comes in. The bicycle is powered by a neuromorphic chip, a chip that could be considered as an artificial brain. It contains many tiny units called memristors that are the equivalent of a biological neuron.
Unlike the CPU, the memristors don’t have the power to perform a lot of functions. But it has enough to perform the function of a single neuron.
The power in neuromorphic computing comes when all these memristors come together, the way that billions of real neurons are connected through synapses. To replicate this, the neuromorphic chip contains arrays of memristors.
The self-driving bicycle has shown us how efficient this new architecture of computing really is. It has 40,000 artificial neurons and 10 million synapses. Compared to an equivalent computer with a standard CPU, the bike can perform up to 100 times faster and uses up to 10,000 times less power, depending on the function it’s performing.
Whilst this sounds extremely promising for the future of AGI, it’s going to be a while until we can fully emulate the human brain. We would need to create a system with 100 billion artificial neurons to achieve this. But hey, we’ve made a bike that has the same number of neurons as a fish, and we have this list of animals by the number of neurons to keep track of our progress!
Maybe we could try for a cockroach next?