The phenomenal binding problem for neural networks

Our aim is to explore neural network mechanisms for phenomenal binding, i.e. combining micro-units of information into the macro-scale conscious experience common in human phenomenology. Such experiential complexity is a key feature that aspiring theories of phenomenal consciousness must account for. We motivate phenomenal binding in a way that aids translation to computational neuroscience, connecting it to related but distinct topics: functional binding, the hard problem of consciousness, and unity of consciousness. We define a deliberately simple artificial neural network (ANN) model, in order to explore its full space of options for implementing phenomenal binding. We demonstrate that the model can implement functional binding but fails to implement phenomenal binding while also maintaining key distinctions between unconscious and conscious processing. We use this set-up to structure possible solutions to p-binding based on which parts of the model they elaborate or which parts of the problem they reject. Several established theories of consciousness map onto our solution structure, such as the aggregation of nodes into complexes applied by Integrated Information Theory (IIT), entanglement collapse in Orch-OR, or the exploitation of field structures in Conscious Electromagnetic Information Theory (CEMI). We also discuss possible solutions open to other theories, such as Global Neuronal Workspace Theory (GNWT) and Dendritic Integration Theory (DIT). Nonetheless, at present, each solution route needs further work, identifying opportunities for researchers to enrich existing theories to account properly for phenomenal binding.