An Information Theory of Consciousness and Motivation

Source: Deep Learning on Medium

An Information Theory of Consciousness and Motivation

Photo by Abishek on Unsplash

The folks at ARAYA in Japan have published a new theory of consciousness called Information Closure Theory (ICT) ( see: https://arxiv.org/abs/1909.13045 ). What leaps out to me about this theory is the use of information loops or closures. I’ve described this before in relationship to Deep Learning architectures that have an uncanny resemblance to Douglas Hoftadter’s Strange Loop:

The second compelling aspect about ICT is that it claims: “ICT demonstrates that information can be the common language between consciousness and physical reality.” This aligns with my explorations on a Theory of Information Emergence (TIE):

I will explore this paper in detail and see where it’s ideas align with my own. ICT claims that every process with a positive non-trivial information closure (NTIC) has consciousness. Furthermore, the level of consciousness is related to the degree of NTIC. ICT further notes that humans do not consciously access information processing at every scale. So for example, we aren’t aware of populations neurons firing and we also aren’t aware of groups of agents coordinating.

ITC defines a closed system where information is decoupled from its environment. From this, it defines further a non-trivial closure (NTC) where a system can internalise and synchronise with the dynamics of its environment. That is, create and maintain its own model of the environment. The degree of NTIC relates to having high predictive power about the environment. These NTC are a macroscopic process that is formed by coarse-graining of the processes of neurons. Only the processes at a certain level of coarse-graining can form a high degree of NTIC. So human consciousness occurs only at the level of coarse-graining where NTIC is highest.

So in this theory, a neural system consists of NTC processes at different scales. We can make the connection with Constraint-closures that I’ve explored previously. The characteristic of NTC is that these are not only self-sustaining but also continuously modeling the environment. Said differently, these processes are generative processes that build accurate models of the environment in a stable and repeatable manner.

I would further add that NTCs are what I call self-models. Self-models are constraint closures that model the ongoing needs of an organism. For humans, there are five self models: bodily, perspectival, volitional, narrative and social. These are all independently constraint closures that model different aspects of self. Humans are conscious of each of these selves, but we see them as a single whole.

The primary difference between ICT and my theory of information emergence (TIE) is that ICT is derived from mutual information principles and TIE is based on constraint-closure and semiotics. Nevertheless, ICT remains a very informative theory that I am certainly looking forward to its further improvement.

However, a theory of consciousness seems to me to be inadequate without exploring the idea of intrinsic motivation. Intrinsic motivation is extremely important because constraint-closures are active agents that are without teachers. Supervised learning and reinforcement learning both require a teacher. Labels for the former and rewards for the latter. What is needed for continual learning is intrinsic motivation. That is, in ICT terms, a preference for building accurate models of the environment. NCT emerges through specific kinds of motivations and these motivations are related to the self-models:

What we need to further elaborate this is a theory of intrinsic motivations. There is a theory called Psi that Joscha Bach has been working on a cognitive architecture:

MicroPsi consists of an orthogonal set of motivations (i.e. needs). To achieve the kind of autonomy we find in human cognition, it is a system not driven by goals. MicroPsi is based on a “minimal orthogonal set of systemic needs” which are signaled to the cognitive system as “urges”. Goals are generated in response to satisfying these needs and avoiding frustration. But needs to constantly change, so the system recalibrates its goals in a dynamical manner. The interesting aspect of the system is that:

“While cognition can change goals, expectations of reward and priorities, it cannot directly influence the needs itself.”