MAKING THE WORLD PREDICTABLE: SOLVING INTRACTABLE INFERENCE THROUGH INFORMATION BOTTLENECKS IN BRAIN, BODY, AND WORLD

Predictive brains are bidirectionally coupled with their environments, in constant interaction to reduce uncertainty by extracting regularities from sensory inputs. Here, we develop a formal information-theoretic approach that links the ideal observer of computational mechanics with the non-ideal observer of cognitive neuroscience, i.e., a Bayesian brain confronted by the intractability problem of performing Bayesian inference. We argue that this intractability is solved through the systematic use of information bottlenecks, which compress and selectively filter environmental information before it is incorporated into internal generative models.
We show that such bottlenecks operate across multiple scales, from the evolution of neural circuits and sensory systems to the active exploratory actions that organisms use in real-time to select and structure their own sensory inputs. A key aspect of our proposal is that information bottlenecks are not only internal and biological, but can also be externalized through actions in the world. By transforming their environments, agents construct external bottlenecks—material and cultural structures that stabilize information, constrain attention, and transmit regularities that are shared across individuals and even across generations. From this perspective, niche construction, material culture, monuments, and even social and cultural prescriptions, can be also understood as external memory devices, collective information bottlenecks that extend the predictive machinery of the brain into the world itself. In sum, the brain solves the intractability problem of inferring information not only by internal neural mechanisms, but by systematically constructing and exploiting information bottlenecks in the world itself.