Eliasmith, C., Stewart, T., Choo, X., Bekolay, T., DeWolf, T., Tang, Y., & Rasmussen, D. (2012). A Large-Scale Model of the Functioning Brain Science, 338 (6111), 1202-1205 DOI: 10.1126/science.1225266
The H. P. Lovecraft novella At the Mountains of Madness is a story about scientific hubris, and the insignificance of human achievement when confronted with the vastness that is the cosmos. The central characters are scientists searching the Antarctic for novel geological or biological forms. They uncover a world of strange, ancient beings, seemingly preserved in the ice. Their discovery has profound implications for the nature of biological evolution, and perhaps the history — and future — of earth itself. The primary entity they find is a god-like tentacled, winged, gilled monstrosity with strange geometrical features. We are given this account:
“It had digestion and circulation, and eliminated waste matter through the reddish tubes of its starfish-shaped base … The nervous system was so complex and highly developed as to leave [the scientist] aghast. Though excessively primitive and archaic in some responses, the thing had a set of ganglial centers and connectives arguing the very extremes of specialized development. Its five-lobed brain was surprisingly advance and there were signs of a sensory equipment, served in part through the wiry cilia of the head, involving facts alien to any other terrestrial organism … It was partly vegetable, but had three-fourths of the essentials of animal structure.”
I had a similar reaction when reading about S.P.A.U.N., the monstrous creation described by Eliasmith et al. in their pape: “A large-scale model of the functioning brain”. The title alone suggests a new and scary precipice of human achievement. And a most sick pape.
S.P.A.U.N. is half way between a robot and a computer program. It has a small camera attached to its head, and a robotic appendage extending from an implied torso that is capable of drawing all manner of digits between 0 and 9. Its brain has multiple sub-systems that independently control the encoding of visual input, reward, working memory, decoding, and motor output. The thrust of Eliasmith et al.’s pape is not that any one component of S.P.A.U.N. is new, but that it can perform not just one task but a variety of tasks, all of which humans can perform, and that its modular architecture resembles and realistically models at least some aspects of the human brain.
How close does S.P.A.U.N. come to resembling the brain? Each of its systems are meant to correspond to cortical and sub-cortical brain areas or functions, though much of the correspondence seems superficial. For example, S.P.A.U.N. has a system for handling “Visual Input”. It’s implied that it corresponds to areas V1, V2, V4, and IT of the primate visual pathway. But S.P.A.U.N. cannot mimic known processing in all those areas because we still don’t know what they do! The visual system of S.P.A.U.N. also reveals that its individual neurons are not as realistic as you might think. The authors stress that it uses biologically-realistic neurons with neurotransmitter dynamics, but most of the visual system instead uses simple linear combinations and thresholding. It’s hard to evaluate these shortcuts because the system is so complex.
So S.P.A.U.N. can do some cool tricks and kind of maybe looks like a brain. Can it fight a bear in one-on-one combat? Definitely not. S.P.A.U.N.‘s proponents admit that it is not the most impressive robot. When I googled “what is the sweetest robot?” I found this strawberry-picking robot. It uses 3D image processing to evaluate ripeness, and can delicately pluck luscious plump strawberries from their stems. That seems more impressive than S.P.A.U.N.
Becoming the best robot is not S.P.A.U.N.’s goal. What is the goal? Building AI systems that perform complex, physical, interactive tasks with flexible, modular systems resembling known biology, that could in principle help us understand how brains work. A longstanding, and recently popularized but criticized, tradition in AI is to build systems that have vast representational power, memory, and fancy statistical learning algorithms, but generally produce simple outputs. For example, a computer vision system that encodes a complex image but only needs to decide which of 10 different kinds of objects it is looking at. But most animals evolved to do more than spit out a 10-bit vector. They engage in real-time with a dynamically changing world. S.P.A.U.N.’s complex motor output and working memory is a step forward, but its emphasis on symbolic reasoning — which number comes next in a sequence? — is excessively abstract. If such symbolic reasoning evolved as an abstraction of more directly-physically-realized functions — when can I jump to make sure I catch the next mouse? — why not first try to build robots that can accomplish those tasks? I recommend the exciting research program of the “embodied robotocist” Rodney Brooks, especially his efforts in the 1990s to build robots with distributed control systems for path finding and navigation. He also appears, looking entirely crazed, in this movie, which features robots, lion taming, tree grooming, and naked mole rats. We are truly standing beneath the Mountains of Madness.