Homeostasis and a Definition of Intelligence
Moravec’s paradox is the observation made by AI researchers that high-level reasoning requires less computation than low-level unconscious cognition. This is an empirical observation that goes against the notion that greater computational capability leads to more intelligent systems. Low level cognition appears to require more computational resources than higher level cognition. This is counter the common intuition that bigger brains leads to higher cognition.
However, we have today computer systems that have super-human symbolic reasoning capabilities. Nobody is going to argue that a man with an abacus, a chess grandmaster or a champion Jeopardy player has any chance at beating a computer. Artificial symbolic reasoning is technology that has been available for decades now and this capability is without argument superior in capability than what any human can provide. Despite this, nobody will claim that computers are conscious.
Today, with the discovery of deep learning (i.e. intuition machines), low-level unconscious cognition is within humanity’s grasp. Let me explore the ramifications of a hypothesis that subjectivity or self-awareness is discovered prior to the discovery of intelligent machines. This is a hypothesis assumes that self-awareness is not a higher reasoning capability.
Let us ask, what if self-aware machines were discovered before intelligent machines. What would the progression of breakthroughs look like? What is the order of these milestones? There is plenty of evidence in nature that simple subjective animals exist without any advanced reasoning capabilities. Let’s assume that it is true that simple subjective machines form the primitive foundations of cognition. How do we build smarter machines from simple subjective machines or machines with simple self models.?
It has been shown in simulations that ego-motion (i.e. bodily self-model) is an emergent property of a curiosity inspired algorithm. This emergence is then followed by the emergence of object detection (i.e. perspectival self-model) and object interaction (i.e volitional self-model). In other words, there is experimental evidence that the foundation to do object detection and interaction is via a self-awareness of where one’s body is situated with respect to space. This then drives the emergence of an agent’s awareness of perspective and then one’s awareness of agency.
The process to achieve ego-motion also allows the reconstruction of 3d space given the image capture of viewpoints. Thus object detection is enhanced in that objects are recognized from different perspectives and objects that occlude one another are identified in their position in 3d space. Furthermore, to achieve 3d interaction with the object, a body needs to know where its articulator is relative to the objects that it can interact with. Therefore, in this experiment, the more computationally demanding task of ego-motion is a requirement to perform a less demanding capability.
The common notion of the progression of intelligence is that higher level cognitive capabilities require more computation. Moravec’s paradox is a tell that this convention is untrue. The quest for consciousness is perhaps the first cognitive capability that needs to be discovered prior to general intelligence. It is incorrect to believe that it is the final goal.
Anil Seth enumerates five different kinds of self-models: bodily, perspectival, volitional, narrative and social selves. These selves are not orthogonal and perhaps partially ordered in what is a prerequisite over another. In his essay “The Real Problem”:
There is the bodily self, which is the experience of being a body and of having a particular body. There is the perspectival self, which is the experience of perceiving the world from a particular first-person point of view. The volitional self involves experiences of intention and of agency — of urges to do this or that, and of being the causes of things that happen. At higher levels, we encounter narrative and social selves.
Anil Seth argues that the problem of understanding consciousness is less mysterious than it is made out to be. There is no hard problem of consciousness.
An AGI roadmap would, therefore, require generating all these five selves in the same order as described above. Autonomy, for example, can be achieved through learning the volitional self. This happens without the narrative self. That is, autonomy is a capability that is discovered prior to the capability of being able to tell stories or participate effectively in a social setting. This clearly makes intuitive sense. To achieve empathic conversational cognition, the narrative and social selves need to be present.
Brendan Lake described an AI roadmap toward “Building Machines that Learn and Think like People.” Lake proposes the development of the following capabilities: (1) build causal models of the world that support explanation and understanding (2) ground learning in intuitive theories of physics and psychology and (3) harness compositionality and learning-to-learn to rapidly acquire and generalize knowledge to new tasks and situations. It is unclear if Lake has prescribed an order on which skill is a prerequisite of which other skill. I propose that we use the notion of self-models to prescribe an ordering for a roadmap for AGI.
Here is the consequence of various selves and the skills that are learned
.
Note: Volitional should be below perspectival in an enactivist framework.
In this formulation, all the world models involve the inclusion of self-models. These self-modes are all ‘Inside Out’ architectures. To understand compositionality, an agent needs an intuitive understanding of the body. To predict physics requires an intuitive awareness of where and what direction one is looking at when one makes an observation. To understand how to learn, one needs to know how to interact with the world.. To understand causality, one needs the capability of following stories. To understand psychology, one needs an understanding of oneself. In summary, you cannot develop any of the skills that Brendan Lake describes without a previous grounding with a model of the self. Self-models are a necessary requirement for the stepping stones of AGI.
AI orthodoxy does not include a notion of self-models in their models. I suspect this is due to either (1) the tradition of science to prefer an objective model of the world or (2) assumption that self-awareness is a higher level of cognitive capability. The second reason creates a bias that research on lower level cognition doesn’t need to take a self-model into account.
A side effect of this model of intelligence based on self-models is that we can also use it to characterize the cognitive capabilities of existing life forms. Below is a radar map that compares an octopus, raven, dog, elephant, human and killer whale across five dimensions that are aligned with a specific self-model.
This diagram is inspired by a research paper “Dimensions of Consciousness” that set out to classify different species according to six dimensions. The six dimensions described in this paper are selfhood, unity, temporality, p-richness (visual), p-richness (touch) and e-richness. The idea is to identify different qualities of consciousness and then identify the extent that a species aligns with these qualities.
I find it more informative to frame consciousness in terms of self-models. The reason for this is that living things are primarily driven by homeostasis. Antonio Damasio in “The Strange Order of Things” argues that the brain’s function at its core is driven by homeostasis. Damasio writes:
Feelings are the mental expressions of homeostasis, acting under the cover of feeling, is the functional thread that links early life-forms to the extraordinary partnership of bodies and nervous systems.
It ensures that life is regulated within a range that is not just compatible with survival but also conducive to flourishing.
One could thus argue that the purpose of brains is the homeostasis of self-models. Thus we are speaking here about homeostasis at an abstract level rather than one that is biological. These self-models afterall are not biologically instantiated, but rather are created in the virtual world of the mind. Anil Seth would call these hallucinations. This is also the nature of an inside-out architecture.
Self-models are homeostatic processes that preserve qualities at different time scales. The agility of an organism to adapt a self-model to a variety of conditions defines a general intelligence in the context of a specific self-model. The natural tension between self-preservation and the need for agility leads to an adaptive cognitive capability. The aggregation of an organism’s self models each responsible for different homeostatic processes and each with a different level of agility corresponds to the holistic general intelligence of the individual.
Species that become dominant due to a few self-model prowess do so at the expense of developing other self-models. As a consequence, cutoff their opportunities toward evolving to a more complex intelligence.
There are many path divergences in human evolution that favored greater agility over more optimal functionality. Human jaws are considerably weaker than the great apes, this deficiency is a consequence of a mutation. However, the consequences of this mutation is that it leads to a more nimble jaw and thus the means for richer vocalization and eventually towards the development of language. The opposable thumb which we share a common ancestor with the great apes. The evolution line of the great apes abandoned this in favor of hands of stronger grip. The weaker human hand however had greater agility to create tools. Weaker hominids were forced to thrive in the savannah instead of the luscious jungles, this led to bipedalism and acute visual perception. Necessity is the mother of invention. Weakness leads to the need for alternative strategies and this can serendipitously lead to greater agility.
The brain is driven by several homeostatic cognitive processes that seek to preserve virtual self-models. The reason why this is so is that the human brain has evolved to decouple functionality from physical implementation. This decoupling is most complete at the highest level of cognition. In other words, we expect to see higher coupling between functionality and physical circuitry in the bodily self-models. However, we should see a complete decoupling in the narrative and in the social self-models. The higher cognitive functions of the mind are software-like. The modular structure of the mini-columns of the isocortex is a tell of this possibility. The physical structure of the isocortex indicates the generality of these components. When we see components with uniform physical characteristics, we can only surmise that the actual computation is performed at an abstract level decoupled from the physical characteristics. It is this decoupling or virtualization of cognition in the human brain that led to its ultimate agility.
Cognition is a constraint satisfaction problem that involves the self, its context, and its goals. In fact, the distinction of inference from learning is likely a flawed bias. Inference happens to also be the same as learning, and both are constraint satisfaction problems. A self-model is what provides meaning, relevance makes the context explicit and agency motivates goals.
An important point here is that the self, the context and the goal are all mental models. Although they may have corresponding real analogs, constraint satisfaction is achieved only with the approximate mental models that are ‘hallucinated’ by cognitive agents. These models are also not static and like a brain in a vat. Rather, agents are decoupled to their environments, agents change with their interaction with the environment.
There is perhaps a synergy between the different selves such that some are prerequisites for others. The reason why the order of skills is extremely important is that the more abstract levels must have the grounding that can be found only in the lower levels. Furthermore, skills that are assumed to be context-free cannot be independent of the context of the self. If we assume Moravec’s paradox to be correct across all cognitive levels, then it is the unconscious bodily self-model (the instinctive level) that requires the greatest computational resources. This implies that contrary to popular consensus, it takes fewer and fewer resources as you move up cognitive levels. Said differently, it takes less and less effort to make exponential progress. This conclusion is very different from the more popular notion that it takes more and more computation to achieve artificial intelligence.
The reason that Anil Seth believes in the insurmountable problem of achieving AGI is that creating an artificial bodily self-model may be too difficult. He writes:
“We are biological, flesh-and-blood animals whose conscious experiences are shaped at all levels by the biological mechanisms that keep us alive. Just making computers smarter is not going to make them sentient.”
The biggest hurdle is at the beginning (i.e. the bodily self-model). This kind of automation simply does not exist in today’s technology. So the acceleration only happens when this capability is achieved, meanwhile, AI innovation has been predominantly determined by the exploitation of brute force computational resources.
I agree with Seth that more computer resources may not ignite into general intelligence. However, I hypothesize the plausibility that simple self-model machines may be that foundation that gets you to general intelligence. The uneasy reality is that this looks very much like a slippery slope. Creating bodily self-models and you just easily can slip into a ditch where you accidentally discover AGI.
An “inside out” architecture is key because it is what we have in the neocortex (aka isocortex). Today’s deep learning is more analogous to insect-like intelligence which is just capable of stimulus-response behavior.
Can we build narrow slices of this cognitive stack and have the stack broaden out over time? For example, a bodily self-model that does not embody the entire sensor network that a human will have. Can we avoid an all or nothing situation and build this incrementally?
The rough sketch for this is that the bodily self-model is developed by learning as an embodied entity in a simulated virtual world. It would have a subset of sensors that is proportional to what can be simulated in this virtual world. The objective is to learn the three lower-level self models (i.e. bodily, perceptive and agency).
These simulations are already being performed today. Once you see this develop with higher fidelity, then I think you’ll see a more rapid acceleration towards general intelligence. There is a tipping point here and that tipping point may be much closer than anyone may have imagined!
More of this here:
