Intelligence from Learnable Novelty

Intelligence appears under different names in different fields: as data compression in statistics and machine learning, as universal computation in dynamical systems, and as adaptive behavior in agents. Each field carries its own objective, and the two most influential drives often fail in mirror image: novelty search, which seeks surprise, is transfixed by a noisy television screen, while the free-energy principle, which avoids surprise, is most content in a dark room. Both failures have a single cause: each objective treats as one quantity the surprise a learner can convert into knowledge and the surprise it never can. Here we show that the learnable part of that information, which we call learnable novelty, yields the seemingly disparate projections of intelligence, and we give a closedform estimator of it built on a cheap and differentiable reservoir computer. Used as a measure, with no supervision of any kind, the estimator recovers decades of complexity classification, ranking the Turing-complete rule 110 highest among the elementary cellular automata.

Introduction. Few concepts are invoked across as many disciplines as intelligence, and few are theorized in as many incompatible ways. To statistics and machine learning it is extreme compression of data; to the study of complex systems it is the emergence of universal computation; in the interaction of an agent with its environment it is open-ended adaptive behavior. Each appearance has its own literature and its own objective function, and the literatures rarely meet. Here we show that these appearances follow from a single principle: the pursuit of learnable novelty. Many creative processes unfold without a preset destination. Biological evolution has no fixed target, and scientific discovery often proceeds without knowing where it will lead. Both illustrate search in which the next direction cannot be specified in advance. Lehman and Stanley [2011] made this intuition operational as novelty search, which abandons objectives, rewards only behavior not seen before, and thereby escapes the deceptive local optima that trap goaldirected search.

Discussion / Conclusion. Complexity generation, abstraction, and exploration are ordinarily studied in separate fields and driven by objectives that owe nothing to one another. In the experiments reported here, all three were produced by a single quantity evaluated by a fixed observer of a single construction. Read as a measure, it recovered the classical complexity ordering of the elementary cellular automata without supervision, placing the one rule proven Turing-complete at the top of the space. Ascended as an objective, its gradient carried a neural cellular automaton from simple dynamics into a regime of complex solitons and organized the representation of an image encoder around the digit classes of MNIST, although no label ever entered training. Handed to an agent as an intrinsic reward, it supplied the exploration that task rewards lack, improving on the task reward in nine of ten environments and collapsing in none. The three phenomena, these results suggest, were never independent: they are projections of one quantity, learnable novelty, onto dynamics, representations, and behavior.