Thinking outside the brain: problem-solving, hardware evolution and embodied cognition

In his 1999 book, the geneticist Steve Jones describes a problem with industrial soap production that he encountered as a young man. The process involved blowing liquid through a nozzle at high pressure in order to produce soap powder. Due to nozzle’s imperfect shape, it was prone to clogging and a better design was sought; however, designing a three dimensional shape that would work more efficiently and consistently than the original was beyond the skills of the engineers, such are the challenges of fluid dynamics. Despite this, a better nozzle was eventually created and the method employed in finding its shape neatly encapsulates the process of evolutionary problem solving:

Take a nozzle that works quite well and make copies, each changed at random. Test them for how well they make powder. Then, impose a struggle for existence by insisting that not all can survive. Many of the altered devices are no better (or worse) than the parental form. They are discarded but the few able to do a superior job are allowed to reproduce and are copied – but again not perfectly. As generations pass there emerges, as if by magic, a new and efficient pipe of complex and unexpected shape (Jones, 1999, pp. 74-75).

I love the nozzle example. It shows how the mindless mechanism behind the evolution of organisms can be harnessed to solve any problem for which we have a clear idea of what ‘good’ looks like. I also like how unusual it is from a biological perspective, with no distinction between genotype and phenotype (in a sense, the shape of the nozzle acts as both). However, the lack of formality in the method leaves much of the process hidden from view. Adrian Thompson’s (1997) study of hardware evolution is a more formal example, with a clear distinction drawn between a software genotype and hardware phenotype, and it is, if anything, even more magical. I don’t think there is a single academic paper that has so completely changed my perspective on science, evolution and the brain.

Thompson’s study was an attempt at using selection to find an electronic circuit capable of discriminating between two tones. He isolated a 10 by 10 corner of a reconfigurable chip, such that the behaviour of just 100 of the chip’s 4096 cells was assessed. A genetic algorithm was then used to create different arrangements of the connections between these 100 cells over successive generations. Specifically, a population of 50 individual arrangements existed in each generation and the relative contribution of any one of these to the next generation was dependent on the extent to which it succeeded in discriminating the tones. After around four thousand generations, an arrangement was found that could discriminate consistently.

The structure of the successful circuit is quite astonishing. A schematic of its arrangement shows that just 21 cells were required to carry out the discrimination and, of these, 5 were special. Like the other 16 they were necessary to ensure that the circuit performed normally, but, bizarrely, there was “no connected path by which they could influence the output” (p.399). In other words, they were contributing to the performance of the circuit in some way other than the direct connections between cells. The means by which these 5 cells exerted their effect isn’t totally clear, but Thompson speculates that it was an aspect of their analogue output such as radiative coupling or temperature modulation. This is supported by the fact that the chip’s performance was sensitive to ambient temperature, working best in the conditions experienced during its evolution.

Thompson’s study vividly demonstrates how evolution is capable of harnessing properties of systems that we are barely even aware of, let alone in a position to fully understand and exploit. Indeed, the physical properties utilised in this case are so peripheral to how we normally understand the function of logic circuits, that its behaviour wouldn’t be captured in a typical simulation of the system. The implications of this for our understanding of how the brain works are, in my opinion, profound. From one point of view Thompson’s study could be viewed as a doctrine of despair: if it’s difficult to understand the function of 100 cells on a chip, how can we hope to achieve even the most basic understanding of brain function? But there’s no need for pessimism. If our progress in understanding the brain has been somewhat disappointing over the past 50 years, this might owe something to our reliance (either explicitly or implicitly) on computer metaphors. Thompson’s findings suggest that we would do well to free ourselves of these constraints, because even if we don’t, the brain already has. In essence this means we need to pay more attention to the environment in which the organism’s behaviour is played out, something AI researchers have learned from the limitations of their simulations. As Dennett (1994) puts it:

Unless you saddle yourself with all the problems of making a concrete agent take care of itself in the real world, you will tend to overlook, underestimate, or misconstrue the deepest problems of design (1994, p.143).

We can already see some success from this approach in the areas of embodied cognition and robotics. Take bipedal walking for example. The old fashioned brute force approach to reproducing this in robots has proved to be extremely challenging and it was thought that humans must be achieving the feat through the employment of considerable computational muscle. This is exemplified to some extent by ASIMO, Honda’s attempt at building a humanoid robot. The project is a couple of decades old now and has cost millions of dollars. The finished product achieves the feat of walking (and myriad other movements) by using a powerful computer to plan the trajectory of its limbs through space. Take a look at a video on the website; Asimo’s movements are a curious mixture of natural and artificial.

It’s not until we compare ASIMO to a simpler walking system that we notice the huge advantages to be gained by understanding how evolved systems exploit their physical shape and surroundings to solve problems. Cornell University’s simple biped, for example, walks by combining the potential energy of being on a hill with the constraints of its weight and form to produce the kind of efficient movement we see in human walkers but not in ASIMO. The principle being exploited is known as passive-dynamics and it refers to the fact that the very shape of the machine predisposes it to moving in a particular way. If you push the biped down a hill it cannot help but walk; walking is just the name we give to the way this type of object falls down hills. In the words of Steve Collins (one of the project’s researchers) ASIMO “is a marvel of engineering but does not utilize the passive-dynamics of its own limbs” consequently ASIMO is far less efficient than a human walker and requires far more computation to achieve the equivalent amount of movement. Have a look at the videos on the Cornell website. The biped clearly lacks the range of movements displayed by ASIMO but it achieves this movement with no computation whatsoever and its walking is undeniably naturalistic; it has a lovely swagger in my opinion. For more evidence of the benefits of exploiting body shape, take a look at the incredible work of the researchers at Boston Dynamics.

It feels like we’re only just beginning to unlock the secrets of how evolved systems achieve goals in ways that are so much less brittle and energetically demanding than traditional engineering and AI approaches. Perhaps we also now have the very beginnings of a new and more satisfying theoretical framework for psychology based on the approach of new fields such as embodied cognition. What form this will ultimately take is not yet clear, but, to paraphrase Richard Dawkins, the new approach will certainly “cause us to think of testable hypotheses that we would otherwise never have dreamed of” (1982, p.2).

References #

Dawkins, R. (1982). The Extended Phenotype. Oxford: W.H. Freeman and Company Limited.

Dennett, D. C. (1994). The Practical Requirements for Making a Conscious Robot. Philosophical Transactions of the Royal Society, 349A, 133-146.

Jones, S. (1999). Almost like a whale: the origin of species updated. London: Random House.

Thompson, A. (1997). An evolved circuit, intrinsic in silicon, entwined with physics. Evolvable Systems: From Biology to Hardware, 1259, 390-405.


Now read this

Initial Conditions

What John Conway’s ‘life’ teaches us about data compression, gene expression and the description of organisms The mathematician John Conway is responsible for a simple but thought-provoking game called ‘life’ that has become a popular... Continue →