Intelligence from Learnable Novelty
18 hours ago
- Intelligence appears in different fields as data compression, universal computation, or adaptive behavior, each with its own objective.
- Two influential drives—novelty search (seeking surprise) and the free-energy principle (avoiding surprise)—fail in mirror image ways, such as being transfixed by a noisy television or content in a dark room.
- Both failures stem from treating the surprise a learner can convert into knowledge and the surprise it cannot as the same quantity.
- The learnable part of information, called learnable novelty, yields the disparate projections of intelligence, and a closed-form estimator is built on a cheap, differentiable reservoir computer.
- As a measure with no supervision, the estimator recovers decades of complexity classification, ranking Turing-complete rule 110 highest among elementary cellular automata.
- As an objective, its gradient guides a neural cellular automaton from simple dynamics into soliton regimes, organizing image encoder representations around MNIST digit classes fully unsupervised.
- As an intrinsic reward for reinforcement learning, it supplies exploration missing from task rewards, improving on baselines in nine of ten environments without collapsing.
- Complexity generation, abstraction, and exploration emerge from ascent on one differentiable quantity, unifying the projections of intelligence on a common quantitative footing.