The human brain is the only working example of general intelligence available. If our goal is to achieve significant progress with AI – it seems to be wise to try to understand what it does, but not necessarily exactly how. There are multiple reasons behind such an approach.
First, it is very complex – as a biological organ that needs to sustain itself, it does much more than it is necessary to process information. There is no need to understand the details of each biochemical reaction – when we would recognize what is the purpose behind it.
Second, no matter how painful it could be – we should once again understand that we are not the center of the Universe – and in the same way, not the ultimate possible intelligence. And the design of our minds has many flaws – which are the source of many personal and collective problems we all experience.
In addition, it is worth noting – that as a civilization we collectively created much more incorrect knowledge than the right one. Just to fix it later…
By taking that (a bit negative) point of view it is easier to realize that in the next decades our current ML/AI solutions will look like primitive toys. They will be improved, again and again, reaching the level of the human brain and ultimately surpassing it.
No matter who will make it – and what will be the role of our company – it is very probable that sooner or later it will happen.
The future of AI will be very different
Room-sized computers of the past and present cloud servers dedicated to AI
The data centers filled with servers will be as exciting as the room-sized computers of the past. And the current ways to interact with AI solutions as popular as command-line interfaces compared to the visual environments of the computer operating systems.
The whole world will interact with AI through multiple channels – visual user interfaces, natural language (voice) and command lines – depending on the context, needs and technical proficiency.
The reason why it will happen will be the progress in AI technology that will become transparent (explainable), energy-efficient (able to be used in more situations), much more capable and applicable for general purposes as we will want to expand the set of usage scenarios and the number of people able to use it.
Such qualities are not in the reach of the current technology, because of its attributes. I sometimes hear that:
„the scientific community accepted that AI can’t be explained”.
Aren’t neuroscientists able to define the regions of the brain responsible for processing specific types of information? As far as I know – we not only share the location of specific regions but also a similar semantic map (location of the word meaning representation) in the brain. For example, verbs are encoded closer to the motor regions and nouns to those responsible for visual perception.
Brains are not random structures. There is a high degree of order, we understand more with each passing decade. It may be said that in the brain there is a high-level structure with low-level adaptation between the individual cells.
This is where neuroscience brings guidance to what is potentially possible with AI when we will focus more on what could be than what is now.
Because if you try to gradually improve something – you may hit a dead end. But if you will define where you want to be in the future – by going backward you may realize what are the crucial points of your plan that just need to be realized in a specific way.
If you are not aiming from the very beginning to create a digital equivalent of a brain – you may not succeed with the goal of building advanced Artificial Intelligence.
The digital equivalent of a human brain
As already stated, the brain is not a random structure optimized by mathematical techniques. It grows on the scaffolding directed by radial progenitor glial cells and our genes. Everything has its place, that is why the location of each brain region is the same in most humans (besides those with some kind of injury).
Learning takes place incrementally. New knowledge, does not disrupt previous if it is still valid and useful. Our own experiences re-shape and adapt low-level representations of the information – this is what makes us unique and better adapted to our own lives.
But it is not as chaotic as in the case of current Deep Networks, described as black-boxes.
The brain exactly knows where to route the specific information coming from the senses. It uses the organ named thalamus to do that. And when we want to remind ourselves about some event – the hippocampus retrieves the binding of multiple data pieces to bring them to our attention.
It does much more than the associative mapping of the input and the output. And uses many more elements to achieve autonomy.
AI is also weaker than small children or pets at many tasks.
The reason for this situation is simple. What makes all living creatures autonomous (in learning and operation) is not the cerebral cortex we associate intelligence with, but all the subcortical components that manage the body, our instincts and unconscious procedures.
The future of Artificial Intelligence
Truly capable and powerful AI needs not only advanced and more complex neural networks, but also a whole set of additional modules.
From this perspective, it may look like we need at least 50-100 years to reach the AGI level. But not for someone who first spent a lot of time to make a good plan about how to reach the ultimate goal – where to look for answers and where not.
Can Deep Learning solve the AGI problem with such energy consumption? Compared to the brain that is multiple times more efficient?
Now, what about datasets collecting and labeling? Do we as humans or animals need thousands of examples to learn anything?
How about the ability to learn new things? With Deep Learning you should start over from scratch each time you need to update the model with the freshly acquired information.
Can something like this become the ultimate invention of human civilization? Or take over the planet as some people suggest?
The brain manipulates both sub-symbolic (sensory perceptions) data and their symbolic generalizations (linguistic or audiovisual labels). This is why:
neuro-symbolic architectures, combining both types are the answer to advanced Artificial Intelligence.
In the brain, data flows in multiple directions not only from left to right or right to left. Why not experiment with 3D neural networks then? Those that are explainable because of the high-level structure and still powerful due to low-level data adaptation. With the ability to be connected to any other computational unit directly or by a set of proxies.
Why not utilize only the most useful parts of extremely large architecture to save time and energy while gaining the benefits of an enormous number of available neurons? If you will eliminate chaos and randomness, you can introduce small changes wherever they are necessary – and make incremental learning not only possible, but very efficient.
Why not equip the AI with a set of numeric parameters simulating a rich set of neurochemicals that drive our autonomy and support making decisions?
To allow the machine to decide which actions are currently prioritized and which should be skipped?
All solutions of the current era (including these currently produced in our company) will be simple and primitive when compared to what will be achieved in the future.
This journey is about defining the right goal. And pursuing it, regardless of required time and effort.
After acquiring enough technical and scientific knowledge to see the clear path to the destination – I am sure that pure neural networks are not enough to deliver truly powerful intelligence.
And by spending years on experimentation, attempts to find the optimal methods I am even more confident that taking more inspiration from neuroscience – will give positive results that not many expected.
We need much more complex and powerful architectures. Such knowledge is both intimidating and empowering. Because after its realization – one can finally move in the right direction.