The Complete Roderick (2005): Why Can’t a Robot Learn Like a Baby?
Recommendation: Read The Complete Roderick and then read Josh Tenenbaum’s work in really making this happen.
An eternal question in artificial intelligence for robotics, or at least since Alan Turing’s time, is Why can’t a robot learn like a baby? The Complete Roderick, a dark screwball comedy sort of a Douglas Adams mashup with Dark Mirror, that addresses this question head on- it assumes that a robot can learn like a baby but then goes further and asks what exactly would a robot learn if it was raised like a child and sent to school: That other children are bullies? That teachers are fallible? That corporations will try to steal and reverse technology? That people act in their own narrow self-interest? That love is often unrequited? Ooops, maybe that’s not what we meant by “learn like a child.”
Roderick is a sentient robot, created in a backwater university, and possessing a sweet childlike innocence. The government never expected the project to be successful; it existed solely to hide millions of dollars being shuffled to other projects. When the government realizes that it may have successfully created a truly intelligent robot, it moves in to shut the robot down for fear that it could be a disruptive technology. To make matters worse, Roderick simultaneously becomes a pawn in a university-corporate partnership scheme where university R&D is deliberately distorted and undervalued in order to make well-connected, rich investors richer. (Yeah, that never happens in real life.) Naturally, the least talented researcher becomes the head of the corporate research project. (Ditto.) The lead scientist rescues Roderick, takes him home to his rather unconventional parents, and the robot boy begins a life among other humans. The other children and teachers refuse to accept him as a robot and treat him as a rather dim child with prostheses. Along the way, the book throws in surprisingly accurate hard science discussions of the works of Turing, Shannon, Wiener, and pretty much everyone involved in the early days of artificial intelligence, as well as a snarky commentary on Shakey AND the Three Laws of Robotics.
Sounds like a satire of our times. Except it was published in 1980 and 1983 as two separate novels (Roderick, or The Education of a Young Machine and Roderick at Random, or the Further Education of a Young Machine) which were reissued as a single novel in 2005.
As Sladek knew, the idea of creating an artificial intelligence that would start with the blank slate of a baby and learn through discovery is as old as the idea of artificial intelligence itself. After all both concepts were introduced by Benedict Cumberbatch, oh I meant Alan Turing, in his timeless 1950 paper “Computing Machinery and Intelligence” that also introduced the Turing Test. The advantages of a Turing "child machine" would be that it should be easy to design the blank slate framework and then the machine could be presented with what it needed to know. The computer would do the work of creating knowledge representations and associations.
As time has gone on, the idea of programming a robot like the tabula rasa of baby has become a major area of serious research. Such an approach could overcome the current barriers in artificial intelligence. For example, babies and children learn deep semantic understanding of language, whereas the current level of natural language is playing a game of probabilities as to what words are likely be spoken. Children learn naive physics, like gravity and friction, long before they can comprehend mathematical equations. Children learn continuously, or what researchers call online learning, while in real-life online learning and robots that may function for years or decades rather than for the span of a lab demo is only now being discussed.
Perhaps one of the most visible efforts in duplicating learning in children is the project led by Dr. Josh Tenenbaum at MIT. A good overview of his group’s work is at MIT Technology Review (www.technologyreview.com). I recently heard him give an engrossing talk at the DARPA AI Colloquium (sort of the defense department version of TED talks). One interesting insight that seems to be producing productive results for computers is that perhaps what children are doing is running a computer game simulation in their heads. For example, I would imagine what will happen if I put a block on the top of another block, first by mentally simulating what will happen (it is not directly on top so it will fall over), then trying it out with real blocks and seeing if it did happen that way, and changing the simulation if necessary. Another insight is that dreams may be a form of adversarial learning: the equivalent of taking “mental images” and then distorting them a bit and rerunning through learning algorithm to train the classifier with more data to better distinguish what an object is or create a better policy for acting.
Several recent robot novels, The Robots of Gotham and Embers of War come to mind, assume “AI nurseries” where artificial intelligence software agents and robots spend years learning and growing. These nurseries are often mentioned as being protected, with the implication that someone or agency might destroy them, but also possibly just to influence or warp their development. In real life there is already a fear over commonly used machine learning datasets being subtly poisoned. The AI nursery meme owes a big debt to Sladek — and if they turn out to be prescient-- to researchers like Tenenbaum.
The Complete Roderick is worth reading, if only for the frisson of how Sladek has essentially written a history of artificial intelligence for robotics. It’s like Jo Walton’s brilliant paean Among Others only with AI researchers substituting for science fiction authors. Readers may find the madcap satire a bit too silly in places, and too dark in others, to be the science fiction equivalent of Catch-22 or Gulliver’s Travels, but it is worth plowing through. The real question raised by The Complete Roderick is not Why can’t a robot learn like a baby? but Would we really want a robot to learn what we learned in school and as a young adult?