The information generation theory of consciousness

James of Seattle called my attention to an interesting paper in the Neuroscience of Consciousness journal: Information generation as a functional basis of consciousness:

  • Drawing upon empirical research into consciousness, we propose a hypothesis that a function of consciousness is to internally generate counterfactual representations detached from the current sensory events.

  • Interactions with generated representations allow an agent to perform a variety of non-reflexive behaviors associated with consciousness such as cognitive functions enabled by consciousness such as intention, imagination, planning, short-term memory, attention, curiosity, and creativity.

  • Applying the predictive coding framework, we propose that information generation is performed by top-down predictions in the brain.

  • The hypothesis suggests that consciousness emerged in evolution when organisms gained the ability to perform internal simulations using generative models.

The theory described broadly matches an idea I’ve pondered several times, that consciousness is about planning, that is, about performing simulations of action-sensory scenarios, enabling non-reflexive behavior.  Although the authors situate their ideas within the frameworks put forth by global workspace and predictive coding theories.

But their theory is more specifically about the internal generation of content, content that might be a prediction about current sensory input, or that might be counter-factual, that is, content not currently being sensed.  In many ways this is similar to the predictive coding framework, but it’s not identical.

In the brain, sensory information flows into early sensory processing regions, and proceeds up through neural processing layers into higher order regions.  But this encoding stage is only a small fraction of the processing happening in these regions.  Most of it is feedback, top down decoding processing from the higher order areas back down to the early sensory regions.

In predictive coding theory, this feedback propagation is the prediction about what is being sensed.  And the feedforward portion is actually error correction to the prediction.  The idea being that, early in life, most of what we get is error correction, until the models get built up, and gradually as we mature, the predictions becomes more dominant.

Importantly, when we’re imagining something, that imagined counterfactual content is all feedback propagation, since what’s being imagined generally has little to no relation to sensory data coming in.  Imagination is less vivid than perceiving current sensations because the error correction doesn’t reinforce the imagery.  (Interestingly, the authors argue that imagery in dreams are more vivid because there’s no sensory error correction to dilute the imagery.)

The information generation theory is that this prediction feedback is what gives rise to conscious experience.  This theory could be seen as similar to recurrent processing theories, although the authors seem to deliberately distance themselves from such thinking by making their point with a non-recurrent example, specifically splitting the encoding and decoding stages into two feedforward only networks.

The authors note that there is a strong and weak version to their hypothesis.  The weaker version is that this kind of processing is a necessary component of consciousness, and is therefore an indicator of it.  The stronger version is that this kind of information generation is consciousness.  They argue that further research is necessary to test both versions.

The hypothesis does fit with several other theories (each having their own strong and weak claim).  The authors even try to fit it with integrated information theory, although they admit the details are problematic.

This is an interesting paper and theory.  My initial reaction is that their weaker hypothesis seems far more plausible, although in that stance it could be seen as an elaboration of other theories, albeit one that identifies important causal factors.  The stronger hypothesis, I think, would require substantially more justification as to why that kind of processing, in and of itself, is conscious.

That’s my initial view.  It could change on further reflection.  What do you think?