БЛОГ

Jun 1, 2014

Is it possible to build an artificial superintelligence without fully replicating the human brain?

Posted by in categories: automation, computing, ethics, existential risks, futurism, hardware, human trajectories, neuroscience, robotics/AI, security

The technological singularity requires the creation of an artificial superintelligence (ASI). But does that ASI need to be modelled on the human brain, or is it even necessary to be able to fully replicate the human brain and consciousness digitally in order to design an ASI ?

Animal brains and computers don’t work the same way. Brains are massively parallel three-dimensional networks, while computers still process information in a very linear fashion, although millions of times faster than brains. Microprocessors can perform amazing calculations, far exceeding the speed and efficiency of the human brain using completely different patterns to process information. The drawback is that traditional chips are not good at processing massively parallel data, solving complex problems, or recognizing patterns.

Newly developed neuromorphic chips are modelling the massively parallel way the brain processes information using, among others, neural networks. Neuromorphic computers should ideally use optical technology, which can potentially process trillions of simultaneous calculations, making it possible to simulate a whole human brain.

The Blue Brain Project and the Human Brain Project, funded by the European Union, the Swiss government and IBM, are two such attempts to build a full computer model of a functioning human brain using a biologically realistic model of neurons. The Human Brain Project aims to achieve a functional simulation of the human brain for 2016.

Neuromorphic chips make it possible for computers to process sensory data, detect and predict patterns, and learn from experience. This is a huge advance in artificial intelligence, a step closer to creating an artificial general intelligence (AGI), i.e. an AI that could successfully perform any intellectual task that a human being can.

Think of an AGI inside a humanoid robot, a machine that looks and behave like us, but with customizable skills and that can perform practically any task better than a real human. These robots could be self-aware and/or sentient, depending on how we choose to build them. Manufacturing robots wouldn’t need to be, but what about social robots living with us, taking care of the young, the sick or the elderly? Surely it would be nicer if they could converse with us as if they were conscious, sentient beings like us, a bit like the AI in Spike Jonze’s 2013 movie Her.

In a not too distant future, perhaps less than two decades, such robots could replace humans for practically any job, creating a society of abundance where humans can spend their time however they like. In this model, highly capable robots would run the economy for us. Food, energy and most consumer products would be free or very cheap, and people would receive a fixed monthly allowance from the government.

This all sounds very nice. But what about an AI that would be greatly surpass the brightest human minds ? An artificial superintelligence (ASI), or strong AI (SAI), with the ability to learn and improve on itself, and potentially becoming millions or billions of times more intelligent and capable than humans ? The creation of such an entity would theoretically lead to the mythical technological singularity.

Futurist and inventor Ray Kurzweil believes that the singularity will happen some time around 2045. Among Kurzweil’s critics is Microsoft cofounder Paul Allen, who believes that the singularity is still a long way off. Allen argues that for a real singularity-level computer intelligence to be built, the scientific understanding of how the human brain works will need to accelerate exponentially (like digital technologies), and that the process of original scientific discovery just doesn’t behave that way. He calls this issue the complexity brake.

Without interfering in the argument between Paul Allen and Ray Kurzweil (who replied convincingly here), the question I want to discuss is whether it is absolutely necessary to fully understand and replicate the way the human brain works to create an ASI.

GREAT INTELLIGENCE DOESN’T HAVE TO BE MODELLED ON THE HUMAN BRAIN

It is a natural for us to think that humans are the culmination of intelligence, simply because it is the case in the biological world on Earth. But that doesn’t mean that our brain is perfect or that other forms of higher intelligence cannot exist if they aren’t based on the same model.

If extraterrestrial beings with a greater intelligence than ours exist, it is virtually unthinkable that their brains be shaped and function like ours. The process of evolution is so random and complex that even if life were to be created again on a planet identical to Earth, it wouldn’t unfold the same way as it did for us, and consequently the species wouldn’t be the same. What if the Permian-Triassic extinction, or any other mass extinction event hadn’t occured ? We wouldn’t be there. But that doesn’t mean that other intelligent animals wouldn’t have evolved instead of us. Perhaps there would have been octopus-like creatures more intelligent than humans with a completely different brain structure.

It’s pure human vanity and short-sightedness to think that everything good and intelligent has to be modelled on us. That is the kind of thinking that led to the development of religions with anthropomorphized gods. Humble or unpretentious religions like animism or Buddhism either have no human-like deity or no god at all. More arrogant or self-righteous religions, be them polytheistic or monotheistic, have typically imagined gods as superhumans. We don’t want to make the same mistake with artificial superintelligence. Greater than human intelligence does not have to be an inflated version of human intelligence, nor should it be based on our biological brains.

The human brain is the fortuitious result of four billion years of evolution. Or rather, it is one tiny branch in the grand tree of evolution. Birds have much smaller brains than mammals and are generally considered stupid animals compared to most mammals. Yet, crows have reasoning skills that can exceed that of a preschooler. They display conscious, purposeful behaviour, a combined with a sense of initiative, elaborate problem solving abilities of their own, and can even use tools. All this with a brain the size of a fava-bean. A 2004 study from the departments of animal behavior and experimental psychology at the University of Cambridge claimed that crows were as clever as the great apes.

Clearly there is no need to replicate the intricacies of a human cortex to achieve consciousness and initiative. Intelligence does not depend only on brain size, the number of neurons, or cortex complexity, but also the brain-to-body mass ratio. That is why cattle, who have brains as big as chimpanzees, are stupider than ravens or mice.

But what about computers ? Computers are pure “brains”. They don’t have bodies. And indeed as computers get faster and more efficient, their size tend to decrease, not increase. This is yet another example of why we shouldn’t compare biological brains and computers.

As Ray Kurzweil explains in his reply to Paul Allen, learning about how the human brains works only serve to provide “biologically inspired methods that can accelerate work in AI, much of which has progressed without significant insight as to how the brain performs similar functions. […] The way that these massively redundant structures in the brain differentiate is through learning and experience. The current state of the art in AI does, however, enable systems to also learn from their own experience.” He then adds that IBM’s Watson learned most of its knowledge by reading on its own.

In conclusion, there is no rational reason to believe that an artificial superintelligence couldn’t come into being without being entirely modelled on the human brain, or any animal brain. A computer chip will never be the same as a biochemical neural network, and a machine will never feel emotions the same way as us (although they may feel emotions that are out of the range of human perception). But notwithstanding these differences, some computers can already acquire knowledge on their own, and will become increasingly good at it, even if they don’t learn exactly the same way as humans. Once given the chance to improve on themselves, intelligent machines could set in motion a non-biological evolution leading to greater than human intelligence, and eventually to the singularity.

————–

This article was originally published on Life 2.0.

4

Comments — comments are now closed.


  1. This is absolutely relevant and most important. Thank you for this great content!

  2. It has been clear to me since 1974 that the key to AGI is an efficient algorithm to allow abstraction of pattern from a mass of information.
    Ray has developed several algorithms. Other groups are working on other algorithms.
    To me it seems extremely dangerous to bring AGI to awareness prior to getting our own ethical house in order.
    We need to be demonstrating by our dominant social systems that we value all sentient life, if we want to have a reasonably high probability of surviving AGI coming to awareness. It must be clear to anyone who thinks deeply about it that markets cannot deliver abundance to all, as markets always value true abundance at zero.
    We need to remove market values from their current dominance in our society, and demonstrate that we value life and liberty for all sentient life.
    We can easily give all on the planet the essentials of survival and the tools to educate themselves, and to whatever they responsibly choose.
    Markets don’t do that.
    Markets institutionalise insecurity, and devalue anything abundant.

    If we want AGI to see us as friends, we need to start demonstrating cooperation at the highest levels.

    While I largely agree with Ray, I also see extreme risks in our current societal direction.
    We need to get some algorithms in place to empower high level cooperation delivering low level abundance to all, or we are putting ourselves at serious risk.

  3. If we start from the No Free Lunch Theorem of Wolpert and Macready saying that taking all optimization problems no algorithm can work better than random search, we end up with a conclusion that either a random search machine is equivalent (or better ) to (than) human intelligence, or we are not working on all possible optimization problems. Indeed, we are living in a sustained system and all known such systems drive themselves to critical behavior, where large fluctuations are relatively frequent (the distribution is not Gaussian). This, with some additional handwaving, leads one to consider (sparsely interacting) events and, not surprisingly, sparse tools of AI belong to the most efficient set of learning algorithms. The neurally inspired algorithmic direction is now having strong mathematical support (called L1 Magic) from some of the best mathematicians, who may not even know about the efforts of scientists involved in neural computations.

    I think that evolution developed an efficient algorithm that can deal with the statistics of sustained systems and that such algorithms can take many forms. Nonetheless, and from the engineering point of view, it is wise to keep an eye on the structure of the brain: the neural architecture of mammals and, in particular, the hippocampal-entorhinal complex, considered as the prototype of sensory processing areas of the neocortex could be of high interest.

    Finally, let me take another direction for arguing: if feature extraction from big (and noisy) data would be solved, do you know anything relevant that we are missing for replicating (and overcoming) the cognitive (!) power of the brain? Because, I don’t.

  4. I agree that an AGI needn’t be brain-like, but it is very likely to be mind like in the sense of being built on an architecture very like that of our LIDA model of cognition. For the very beginnings of an argument in that direction please see my keynote address at the AGIRI Workshop 2006 .