May 2016 …. Out of nowhere I snapped awake at 3am with an epiphany about the nature and basis of self. Everything was so clear.
The idea sat with me for a few weeks while I digested it and then I took the plunge and called up one of the local universities “ hey I have this idea that I think can explain consciousness …would you like to collaborate?” There was a long awkward silence followed by “you know we get a lot of calls from people with theories” suffice to say this was not a long call.
Over the years the idea has remained, and I roll it out occasionally at Perth Machine learning group meetings, however recently it seems to have become more of a topic de jour and with rapidly advancing field of AI the elaboration of these ideas now falls well within the range of current capabilities.
The idea stems from the questions: why do we have consciousness? What problem is consciousness trying to solve? what evolutionary advantage does it give? and do we see evolutionary precursors to consciousness?
With these questions swirling around my head, my brain distilled down to the thought that that the answer lay in the differentiation between self and other (non-self). This is an important ability exhibited by most living systems at some level of organisation. Even at the most basic level of me (not food) vs other (food) being able to differentiate is critical to survival.
The simplicity and ubiquity of such a ability suggests a conserved mechanism for a simple classifier for self (a self-classifier). Thinking about the potential of this simple mechanism and how and why organisms might have evolved to use it leads us on an interesting journey.
For bandwidth limited biological systems, with limited energy budgets, being able to do this self-recognition task efficiently would be an important mechanism for simplifying living in complex environments and tracking where signals have come from.
The world is perceived by sensory input. Filtering that input into streams to separate self-associated events from external sources significantly reduce the amount of information you need to attend to in external signals or tracking internal state. I like to picture it like a screen onto which the entire sensory input is projected, the classifier rakes over this and simply tags the streams as internal or external for later processing.
This simple self-classifier provides critical context to environmental inputs, it allows the integration of information, a way to track basic internal state and a memory/ persistence of self.
Further, when learning agency in the world, the first step to learn to interact with it. The differentiation of self-caused from non-self-caused events simplifies the process of causal interpretation and builds an understanding of the causal chain of events. At a basic level being able to recognise that an action was your own action is fundamental to learning about your body and how it interacts with the world. In later life this understanding of the body can be extended to include new concepts or extensions to the body, i.e. tools.
In higher organisms’ Internal mental processes (many of which are generative) are caught by this stream filter and identified as self, for example recognition of internal voice or, if faulty, hearing voices not associated with self.
Thought processes, articulated by internal voice, are most often associated with self, while internally generated visualisation may contain a mixture of self and non self. Overall, this simple filtering of streams gives rise to a strong association of a particular collection of internal processes as self.
Interestingly this concept can also be applied to superorganisms. Think of army ants spreading out over the forest floor, they are motion sensitive and snap at anything that moves and yet there are in a sea of moving ants. At an individual level they differentiate non colony from colony before they lunge and snap when detecting motion.
So we can now imagine how a simple mechanism shared by many forms of life could been reused and form our identity of self and be the basis of our causal understanding of the world.
The takeaway is that all the richness we associate with consciousness is probably derived from a simple mechanism which is shared by most other organisms. We exist not apart from other animals but on a spectrum of ability to articulate our internal processes.
Building AI systems that discriminate their sensory input in this way is not difficult.
Adding tools for provable reasoning, chain of thought, associative memory and theory of mind have also been addressed by multiple researchers. Providing such a model with articulation via a language model is also trivial (not to be confused with just talking to a language model driven chat bot)
The question is are people ready for what can be made?
Writer Sean Driver