On the simulation (and energy costs) of human intelligence, the singularity and simulationism

For many researchers the Holy Grail of robotics and AI is the creation of artificial persons: artefacts with equivalent general competencies as humans. Such artefacts would literally be simulations of humans. Some researchers are motivated by the utility of AGI; others have an almost religious faith in the transhumanist promise of the technological singularity. Others, like myself, are driven only by scientific curiosity. Simulations of intelligence provide us with working models of (elements of) natural intelligence. As Richard Feynman famously said ‘What I cannot create, I do not understand’. Used in this way simulations are like microscopes for the study of intelligence; they are scientific instruments.

Like all scientific instruments simulation needs to be used with great care; simulations need to be calibrated, validated and – most importantly – their limitations understood. Without that understanding any claims to new insights into the nature of intelligence – or for the quality and fidelity of an artificial intelligence as a model of some aspect of natural intelligence – should be regarded with suspicion.

In this essay I have critically reflected on some of the predictions for human-equivalent AI (AGI); the paths to AGI (and especially via artificial evolution); the technological singularity, and the idea that we are ourselves simulations in a simulated universe (simulationism). The quest for human-equivalent AI clearly faces many challenges. One (perhaps stating the obvious) is that it is a very hard problem. Another, as I have argued in this essay, is that the energy costs are likely to limit progress.

However, I believe that the task is made even more difficult for two further reasons. The first is – as hinted above – that we have failed to recognize simulations of intelligence (which all AIs and robots are) as scientific instruments, which need to be designed, operated and results interpreted, with no less care than we would a particle collider or the Hubble telescope.

The second, and more general observation, is that we lack a general (mathematical) theory of intelligence. This lack of theory means that a significant proportion of AI research is not hypothesis  driven, but incrementalist and ad-hoc. Of course such an approach can and is leading to interesting  and (commercially) valuable advances in narrow AI. But without strong theoretical foundations, the grand challenge of human-equivalent AI seems rather like trying to build particle accelerators to understand the nature of matter, without the Standard Model of particle physics.

The text above is the concluding discussion of my essay On the simulation (and energy costs) of human intelligence, the singularity and simulationism, which appears in an edited collection of essays in a book called From Astrophysics to Unconventional Computation. Published in April 2019, the book marks the 60th birthday of astrophysicist, computer scientist and all round genius, Susan Stepney.

Note: regular visitors to the blog will recognise themes covered in several previous blog posts, brought together in I hope a coherent and interesting way.

Estimating the energy cost of evolution

Want to create human-equivalent AI? Well, broadly speaking, there are 3 approaches open to you: design it, reverse-engineer it or evolve it. The third of these - artificial evolution - is attractive because it sidesteps the troublesome problem of having to understand how human intelligence works. It's a black box approach: create the initial conditions then let the blind watchmaker of artificial evolution do the heavy lifting. This approach has some traction. For instance David Chalmers, in his philosophical analysis of the technological singularity, writes "if we produce an AI by artificial evolution, it is likely that soon after we will be able to improve the evolutionary algorithm and extend the evolutionary process, leading to AI+". And since we can already produce simple AI by artificial evolution, then all that's needed is to 'improve the evolutionary algorithm'. Hmm. If only it were that straightforward.

About six months ago I asked myself (and anyone else who would listen): ok, but even if we had the right algorithm, what would be the energy cost of artificially evolving human-equivalent AI? My hunch was that the energy cost would be colossal; so great perhaps as to rule out the evolutionary approach altogether. That thinking, and some research, resulted in me submitting a paper to ALIFE 14. Here is the abstract:

This short discussion paper sets out to explore the question: what is the energy cost of evolving complex artificial life? The paper takes an unconventional approach by first estimating the energy cost of natural evolution and, in particular, the species Homo Sapiens Sapiens. The paper argues that such an estimate has value because it forces us to think about the energy costs of co-evolution, and hence the energy costs of evolving complexity. Furthermore, an analysis of the real energy costs of evolving virtual creatures in a virtual environment, leads the paper to suggest an artificial life equivalent of Kleiber's law - relating neural and synaptic complexity (instead of mass) to computational energy cost (instead of real energy consumption). An underlying motivation for this paper is to counter the view that artificial evolution will facilitate the technological singularity, by arguing that the energy costs are likely to be prohibitively high. The paper concludes by arguing that the huge energy cost is not the only problem. In addition we will require a new approach to artificial evolution in which we construct complex scaffolds of co-evolving artificial creatures and ecosystems.

The full proceedings of ALIFE 14 have now been published online, and my paper Estimating the Energy Cost of (Artificial) Evolution can be downloaded here.

And here's a very short (30 second) video introduction on YouTube:

My conclusion? Well I reckon that the computational energy cost of simulating and fitness testing something with an artificial neural and synaptic complexity equivalent to humans could be around 10^14 KJ, or 0.1 EJ. But evolution requires many generations and many individuals per generation, and - as I argue in the paper - many co-evolving artificial species. Also taking account of the fact that many evolutionary runs will fail (to produce smart AI), the whole process would almost certainly need to be re-run from scratch many times over. If multiplying those population sizes, generations, species and re-runs gives us (very optimistically) a factor of 1,000,000 - then the total energy cost would be 100,000 EJ. In 2010 total human energy use was about 539 EJ. So, artificially evolving human-equivalent AI would need the whole human energy generation output for about 200 years.

---

The full paper reference:

Winfield AFT, Estimating the Energy Cost of (Artificial) Evolution, pp 872-875 in Proceedings of the Fourteenth International Conference on the Synthesis and Simulation of Living Systems, Eds. H Sayama, J Rieffel, S Risi, R Doursat and H Lipson, MIT Press, 2014.

Related blog posts:

Why robots will not be smarter than humans by 2029

How Intelligent are Intelligent Robots?

When giving talks about intelligent robots I've often been faced with the question "how intelligent is your robot?" with a tone of voice that suggests "...and should we be alarmed" It's a good question but one that is extremely difficult - if not impossible - to answer properly. I usually end up giving the rather feeble answer "not very", and I might well add "perhaps about as intelligent as a lobster" (or some other species that my audience will regard as reassuringly not very smart). I'm always left with an uneasy sense that I (and robotics in general) ought to be able to give an answer to this perfectly reasonable question. (Sooner or later I'm going to get caught out when someone follows up with "and exactly how intelligent is a lobster?")

Given that the study of Artificial Intelligence is over 60 years old, and that of embodied AI (i.e. intelligent robotics) not much younger, the fact that roboticists can't properly answer the question "how intelligent are intelligent robots" is, to say the least, embarrassing. It is I believe a problem that needs some serious attention.

Let's look at the question again. There is an implied abbreviation here: what my interlocutor means is: how intelligent are intelligent robots when compared with animals and humans? What's more we all assume a kind of 'scale' of intelligence - with humans (decidedly) at the top and, furthermore, a sense that a crocodile is smarter than a lobster, and a cat smarter than a crocodile. Where, then, would we place a robot vacuum cleaner, for instance, on this scale of animal intelligence?

Ok. To answer the question we clearly need to find a single measure, or test, for intelligence that is general enough it can be applied to robots, animals or humans. It needs to have a single scale broad enough to accommodate human intelligence and simple animals. This metric - let's call it GIQ for General (non-species-specific) Intelligence Quotient - would need to be extensible downwards - to accommodate single celled organisms (or plants for that matter) and of course robots because they're not very smart. Thinking ahead it should also be extensible upwards for super-human AI (which we keep being told is only a few decades away). Does such a measure exist already? No, I don't think it does, but I did come across this news posting on physorg.com a few days ago with the promising opening line How do you use a scientific method to measure the intelligence of a human being, an animal, a machine or an extra-terrestrial? It refers to a paper titled Measuring Universal Intelligence: Towards an Anytime Intelligence Test. I haven't been able to read the paper (it is behind a paywall) but - even from the abstract - it's pretty clear the problem isn't solved. In any event I'm doubtful because the news writeup talks of "interactive exercises in settings with a difficulty level estimated by calculating the so-called Kolmogorov complexity", which suggests a test that the agent being tested has to engage in. Well that's not going to work if you're testing the intelligence of a spider is it?

So let's set aside the problem of comparing the intelligence of robots with animals (or ET) for a moment. Are there existing non-species specific intelligence measures? This interesting essay by Jonathan Ball: The question of animal intelligence outlines several existing measures based on neural physiology. In summary they include:

• Encephalization Quotient (EQ): which measures whether the brain of a given species is bigger or smaller than would be expected, compared with that of other animals its size (winner: Humans)
• Cortical Folding: a measure based on the degree of cortical folding (winner: Dolphins)
• Connectivity: a measure based on comparing the average number of connections per neuron (winner: Humans)

Interestingly, if we take the connectivity measure - which Jonathan Ball suggests offers the greatest degree of correlation with intelligence - then if our robot is controlled by an artificial neural network we might actually have a common basis for comparison of human and robot intelligence.

So, even if none of them are entirely satisfactory it's clear that there has been a great deal of work on measures of animal intelligence. What about the field of robotics - are there intelligence metrics for comparing one robot with another (say a vacuum cleaning robot with a toy robot dog)? As far as I'm aware the answer is a resounding no. (Of course the same is not true in the field of AI where passing the Turing Test has become the iconic - if controversial - holy grail.)

But all of this presupposes, firstly, that we can agree on what we mean by 'intelligence' - which we do not. And secondly, that intelligence is a single thing that any one animal, or robot, can have more or less of* - which is also very doubtful.

---

*An observation made by an anonymous reviewer of one of my papers, for which I am very grateful.

On optimal foraging, cod larvae and robot vacuum cleaners

On Monday I took part in a meeting of the Complex Systems Dynamics (CoSyDy) network in Warwick. The theme of the meeting was Movement in models of mathematical biology, and I heard amazing talks about (modelling) albatross flight patterns, e-coli locomotion, locust swarming and the spread of epidemics. (My contribution was about modelling an artificial system - a robot swarm.) Although a good deal of the maths was beyond me, I was struck by a common theme of our talks that I'll try and articulate in this blog post.

The best place to start is by (badly) paraphrasing a part of Jon Pitchford's brilliant description of optimal foraging strategies for cod larvae. Cod larvae, he explained, feed on patches of plankton. They are also very small and if the sea is turbulent the larvae have no chance of swimming in any given direction (i.e. toward a food patch), so the best course of action is to stop swimming and go where the currents take you. Of course the food patches also get washed around by the current so the odds are good that the food will come to you anyway. There's no point wasting energy chasing a food patch. Only if the sea is calm is it worthwhile for the cod larvae to swim toward a food patch. Thus, swim (toward food) when the sea is calm, but don't swim when it's rough, is the optimal foraging strategy for the cod larvae.

It occurred to me that there's possibly a direct parallel with robot vacuum cleaners, like the Roomba.  A robot vacuum cleaner is also foraging, not for food of course, but dirt in the carpet. For the robot vacuum cleaner the equivalent of a rough, turbulent, sea is a room with chaotically positioned furniture. The robot doesn't need a fancy strategy for covering the floor: it just drives ahead and every time it drives up to a wall or piece of furniture it stops to avoid a collision, makes a random turn and drives off again in a straight line. This is the robot's best strategy for reasonable coverage (and hence cleaning) of the floor in a chaotic environment (i.e. a normal room). Only if the room was relatively large and empty (i.e. a calm sea) would the robot (like the cod larvae) need a more sophisticated strategy for optimal cleaning - such as moving in a pattern across the whole area to try and find all the dirt.

Robot vacuum cleaners, like cod larvae, can exploit the chaos in their environment and hence get away with simple (i.e. stupid) foraging strategies. I can't help wondering - given the apparently unpredictable current economic environment - if there's really no point governments or individuals trying to invent sophisticated economic strategies. Perhaps the optimal response to economic turbulence is the KISS principle.

Why large robot swarms (and maybe also multi-cellular life) need immune systems.

Just gave our talk at DARS 2010, basically challenging the common assumption that swarm robot systems are highly scalable by default. In other words the assumption that if the system works with 10 robots, it will work just as well with 10,000. As I said this morning "sorry guys, that assumption is seriously incorrect. Swarms with as few as 100 robots will almost certainly not work unless we invent an active immune system for the swarm". The problem is that the likelihood that some robots partially fail - in other words fail in such a way as to actually hinder the overall swarm behaviour - quickly increases with swarm size. The only way to deal with this - and hence build large swarms - will be to invent a mechanism that enables good robots to both identify and disable partially failed robots. In other words an immune system.

Actually - and this is the thing I really want to write about here - I think this work hints toward an answer to the question "why do animals need immune systems?". I think it's hugely interesting that evolution had to invent immune systems very early in the history of multi-cellular life. I think the basic reason for this might be the very same reason - outlined above - that we can't scale up from small to huge (or even moderately large) robot swarms without something that looks very much like an immune system. Just like robots, cells can experience partial failures: not enough failure to die, but enough to behave badly - badly enough perhaps to be dangerous to neighbouring cells and the whole organism. If the likelihood of one cell failing in this bad way is constant, then it's self-evident that its much more likely that some will fail in this way in an organism with 10,000 cells than 10 cells. And with 10 million cells (still a small number for animals) it becomes a certainty.

Here is the poster version of our paper.

Symbrion debates @Stuttgart

In Stuttgart, at the University, for a Symbrion project meeting. Its been a really tough meeting - which is hardly surprising given that we're one year in and - next month - have the big end-of-first-year review meeting in Prague. So a major part of the meeting has been a dress rehearsal for the review.

However, spending a day and a half with a group of very smart people is always a pleasure, and there were some really interesting issues to debate. One concerns the fundamental question of how much of the Symbrion system should be designed and how much evolved (using evolutionary computing techniques). One could take a purist view and aim to evolve every aspect. My own view is more pragmatic. I think that achieving the aims of the Symbrion project is going to be so difficult that we should resort to artificial evolution only for the parts of the system that we can't design, because we don't know how.

Also, I think there's a 'biological plausibility' argument for taking the pragmatic view. The Symbrion system will be both a swarm of individual robots, behaving like a swarm, and - following self-assembly - a multi-cellular organism, behaving as a single organism. Swarm and organism have, I think, radically different control paradigms; the former fully decentralised and dependent on mechanisms of emergence and self-organisation, the latter centralised and coordinated (by a central nervous system). Of course ant genes must both contain the instructions to build multi-cellular animals (the ants with CNSs and coordinated control, e.g. for walking), and their behaviours which give rise to the colony's collective swarm intelligence. However, Symbrion goes beyond anything seen in nature. We want the Symbrion robots to sometimes behave like complicated ant-like creatures, and sometimes behave like complicated cells in a complex body (that can perform useful coordinated functions). I think if such a thing were possible to be evolved it would have been (except for the fascinating but much-simpler-than-Symbrion case of the social amoeba Dictyostelium discoideum sometimes self-assembling into multicellular structures).

This is why I think engineering a single evolutionary process that can evolve both swarm intelligent control and centralised coordinated control is asking too much.

Some really interesting Big Questions

As reported in my previous two blog entries How and Why do we have Culture?, and What do Aliens look like?, I've become involved in National Science Week, a major part of which has been structured around Big Questions. I've had some fun the last few days responding to some of the big questions - which are really good. Here are the questions and responses I've made so far:

When will people invent robots to fight in wars?

Could you create a robot with a mind of its own?

Can we invent a robot that will do everything for you?

Will the human brain ever be capable of understanding itself?

What actually is Intelligence? Do we understand it?

Is it possible that AI technology may get out of control?

What do Aliens look like?

Amazing meeting yesterday afternoon at the Science Museum. Here's the story: the British Association for the Advancement of Science (BA) has been asking for science questions via their web pages for awhile, in advance of Science Week, which is 7-16 March. A question that keeps coming up is "what do aliens look like?" so, to address that question, the BA pulled together a small panel which met yesterday. Our brief was to come up with some plausible alien life forms that can be visually presented during Science Week. The keyword here is plausible. It would be easy to pluck super-exotic aliens from the rich fauna of SF but then very difficult to explain the science. Of course, the fictional evolutionary science, or biochemistry, or ecosystems of even plausible aliens is going to have gaping holes, but we were tasked with trying to minimise those.

So, what did we come up with..? I can't say now but all will, I hope, be revealed during Science Week (and in this blog, then). ...and here is the press release describing our life forms.


Visualisation by Julian Hume, Research Fellow at the Natural History Museum/University of Portsmouth.

What I've changed my mind about and Why

Every New Year the Edge poses a question and publishes the responses from among the world's intelligentsia. It's fun to read the responses but, even more fun to write your own.

This year's question is What have you changed your mind about? Why? With the rather important sub-text science is based on evidence. What happens when the data change? How have scientific arguments or findings changed your mind?

The sub-text makes all the difference. Without that we can all think of things we've changed our minds about. But how often are we faced with major paradigm shifts from one set of views to a completely different set, based on scientific evidence, and quite possibly in the face of our own prior beliefs or prejudices. Rarely, I'll bet. Most scientists tend to be educated within a particular tradition and then remain within its established dogma throughout their working lives.

Without doubt the most profound scientific change of mind I have experienced (which took many years) is in relation to Newton's first law of motion, which I now believe to be wrong. However, I discussed this two years ago, in response to the Edge 2006 New Year question, so I won't repeat it here.

So, what else have I changed my mind about?

Well, in November 2007 I was lucky enough to be invited to a University of Bristol workshop on Mathematical Models of Cognitive Behaviour. The assembled speakers covered a broad spectrum of topics, but one in particular really caught my imagination. Dennis Bray's brilliant and illuminating talk introduced the computational processes going on inside the single cell, e-coli. I learned, to my astonishment, that single-celled micro-organisms have a complex repertoire of behaviours including memory and adaptation (in addition to finding food and reproduction). Behaviours that one normally associates with far more complex multi-celled organisms with nervous systems.Yet these are creatures with no nervous system. But, as I learned, an exquisitely complex set of molecular computational processes are present within the dense chemical soup (cytoplasm) encapsulated within the cell. Dennis explained that for e-coli, for instance, we now have a kind of circuit diagram of the complete set of computational processes. But it's far from simple!

Before Dennis' talk I was fond of saying that the robots in our lab are simpler than the simplest animals. As a result of what I learned from his talk, I've now seriously downgraded my assessment:

Our robots are simpler than the simplest single-celled micro-organisms.