Binding problem and integration of information in consciousness

McFadden (2020) takes the definition of binding from Treisman (1999): the capacity to integrate information across time, space, attributes and ideas within a conscious mind. The problem is often posed for vision, where colours, textures, lines and motion are processed in distinct brain regions yet form one percept. McFadden argues that binding is a general feature of consciousness in all its modes¹. He gives a line of Shakespeare as an example, in which four discordant nouns are bound into one insight. He also cites Gary Marcus, who laments that AI lacks this ability¹.

McFadden separates causal integration from physical integration. A logic gate's output bit depends on both input bits, so the gate integrates them causally. The single output bit cannot encode both inputs, however, so it is not physically integrated information². He adds that the inputs are free to accept new data once the signal has passed. No physical state then corresponds to the integration the gate performed².

Friston, Wiese and Hobson (2020) note a gap between the free energy principle and integrated information theory (IIT). The first is framed as probabilistic beliefs about something external. IIT concerns probability distributions over the states of a system³. They nonetheless judge, on an informal review, that the information geometry of Markov blankets appears to meet the qualities that form IIT's axiomatic basis⁴. Fields, Glazebrook and Levin (2021) tie positive integrated information to internal memories and a separable internal state⁵. Bongard and Levin (2023) quote Hoel on the exclusion postulate. Many scales of description carry some integrated information, so the postulate is needed to get a single consciousness, and it is IIT's most controversial postulate⁸.

Several later works place integration outside the brain. McMillen and colleagues (2021) suggest that long-distance communication may force us to re-evaluate cell autonomy, as more evidence emerges that a cell's behaviour integrates information from throughout the body⁶. Fields and Levin (2022) describe interoception models being integrated with perception in a shared memory global workspace architecture⁷. Reid (2023) discusses multi-attribute problems in the slime mould Physarum. Solving them requires comparing options by relative differences and integrating information along several axes, not applying a single threshold⁹.