Consciousness

Consciousness is the presence of phenomenal experience. A system is conscious if there is something it is like to be that system.

A conscious system has direct access only to its own experience. Experience can model things outside, point to concepts, mislead, memories can be false. But the current, transient experience itself is the one tautological fact a conscious being can be sure of. Cogito, ergo sum.

This is where I struggle to understand the “consciousness is an illusion” crowd. The fact you experience in this moment, is all you need to know you are conscious. That its phenomenal content might not be representative of what is the case is irrelevant for this claim. Consciousness is a claim about the presence of experience not its content.

This experiential closure makes consciousness unusually difficult to define. This initial definition is effectively ostensive: we point to experience itself. Another difficulty is that from the perspective of a conscious being nothing is provable, even math which contains absolute truths through strict logical consistency, has to rely on experience for comprehension. But it never was about truth, but predictability. Something is true to the degree it explains and predicts.

So lets find the most predictive theory of consciousness. I adopt computational functionalism as an epistemic foundation. It is the most actionable and rational one I encountered so far. Computational posits anything is made up of discrete state transitions, no infinite resolution, only unboundedness. Functionalism is the position that a thing is exactly defined by its causal structure, internal and external. This distinguishes it from Behaviorism which defines a thing only by its external behavior.

The Hard Problem

The hard problem is about answering the question of why anything feels like something. How does matter become mind?

For materialists this question might seem unanswerable, which is why idealists and panpsychists often take this to support their stance. But idealists and panpsychists do not answer the question either, they sidestep the question by making consciousness an axiom.

I do not have a definite answer to the hard problem, but I would like to point to the confusions that made it appear in the first place. Whatever we call physics is just a model from the perspective and at the resolution of us as observers. The real physical universe is way more abstract, the closer we look, the weirder it gets (time dilation, quantum effects…).

Outside our subjective experience, in the true physical universe, there is no color, no sound, no categories, no objects, no atoms, no space, no time, no you, no I, no mind. These are concepts local to our mind.

But where is our mind then?

The idea is to take computation and causal insulation seriously. Software is a causal pattern running on a computational substrate. Given the right emulator, it can run on any substrate (biological, thermodynamic, quantum, photonic, traditional…). Turing-completeness is not a high bar and entails any software can be run substrate agnostic. On the substrate, it can create arbitrary worlds, independent of the outside universe the insulator exists in. It allows you to nest worlds like matryoshka dolls, each with its own laws of physics.

This is where I suspect some formulations of the hard problem mix levels of description. Matter cannot become conscious. Matter can facilitate computation that creates consciousness inside a simulation. This reframes the problem as identifying which computational organization constitutes experience. It does not yet explain why that organization must be experiential. I think Franz put it really well when saying it is not that patterns feel like something, that they can be experienced, it is that there is a feels-like pattern. I personally like to put it as self-containment, that a quale itself is a piece of self-executing code, instead of being pointer-based, like natural language.

This is not a purely physicalist stance, it is clearly not idealist either, nor is it panpsychist. It is functionalist, assuming only that there is discernible difference (information), forming patterns and one of them, consciousness, is the stream in which experience takes shape as bundles of qualia, feels-like patterns.

I believe the hard problem is often ill-posed in the sense that it is inconsistent, matter cannot become mind, the same way LEDs cannot become Spiderman, it is conflating different levels of existence. A video of spiderman observed by us on the LED screen is not instantiating spiderman in this world. However it can bootstrap spiderman in your mind using your mind as the computational substrate.

But if you would like to create spiderman with LEDs independent of any observer they would need to be arranged as a computer and then run a simulation of the world spiderman lives in, then the spiderman pattern can be placed in this simulated world.

Notice that from the functionalist perspective, this would truly be spiderman. But from our perspective below, at its substrate level, there is no spiderman. We can research the seemingly random flickering of the LEDs and detect at some point that there is a causal pattern which constitutes spiderman in context of the other LEDs flicker patterns which constitute his world. By reasoning through this, we could parse the child universe in the LED interactions but for spiderman to exist in our universe we need a red suit and cobweb spraying hands, not flickering LEDs. Diving into the LED-substrate universe however this is exactly what we would see.

The same goes for consciousness, yes, you cannot imagine how atom interactions can suddenly experience the world and feel like something. But you mistakenly attribute ontological status to your models and perceptions, assuming your world model as ontological physical reality. I am not claiming that we need physics we are not able to see or model, I am simply saying that consciousness exists in a simulation only, it exists one level of abstraction above its substrate. Something cannot feel like anything if it is not on its not implemented on the same layer of existence, a shared layer below needs to nest a world model and code qualia as the relation to this world model. This layer below, the true physical world, is inaccessible experientally.

The reason it is so tricky to grasp and understand is that we are conscious and stuck in not being able to reference anything outside our experience and therefore world model. I speculate, that we mistake our consciousness stems from the physical universe, while it stems from the relation to our world model. Both consciousness and world model have to be on the same level of abstraction inside the a computer, which we refer to as our brain.

But phenomenal experience itself stays private. There is nothing I can know except that I experience now.

Researching consciousness

With those metaphysical cards on the table, researching consciousness has a clearer path forward. We do not need to rely on correlates of consciousness in the human brain, studying noisy signals with crude sensors. Neuroscience faces serious constraints: access to the system, measurement resolution, experimental control, and the difficulty of separating interacting processes.

Mechanistic interpretability is the neuroscience equivalent for digital minds. As we have full control over the hardware and software, we can in theory extract all its information (weights, activations, metadata) and replay any situation. In practice, complete access is not complete understanding. Complexity remains a bottleneck, necessitating its own heuristic tools, like SAEs, J-lens and transcoders.

We can use this access to our advantage. We can test our understanding of consciousness by building a digital consciousness.

This has huge ethical implications, and they go both ways. Not trying to understand consciousness risks creating suffering through ignorance. Trying to understand it by building candidate systems creates a risk of producing suffering during the research itself. It is dangerous to take any unjustified position for or against the possibility of suffering machines. Crude conviction in both ways can cause a lot of harm.

Consciousness as a self-organizing learning algorithm

The hypothesis of consciousness as a self-organizing learning algorithm was first publicized by Joscha as one of the central hypotheses of the CIMC.

I, as a human, experience things, I am conscious. Sometimes I’m not, I sleep without dreaming, I lose consciousness on anesthesia. I am not a solipsist, I believe all humans are conscious by default.

Apparently most people believe that as we empathize with humans and have ethical concerns regarding many life forms. All current life forms are made out of cells. They self-organize into nested control structures and consume negentropy to stay alive. The relationship between biological self-maintenance and consciousness begs the question whether they share organizational principles. The moment of death is the moment the control hierarchy collapses and disintegrates, global coherence is lost.

Why self-organizing?

In self-organizing systems, components apply local rules to adapt their interactions in response to changing conditions and cooperatively realize adaptation. (Danny Weyns)

Biological systems are inherently self-organizing. Anything that runs on an organism with changing network topology requires continuous adaptation and flexibility in its implementation. Early development and recovery from injury illustrate how flexibly neural systems can sometimes allocate functions and compensate for disruption (see 1). If consciousness runs in self-organized systems such as brains, it needs to cope with those changes.

Local coordination can repair without requiring one controller to know every detail, making self-organization quite an efficient paradigm. Additionally, distributed organization can avoid some central points of failure. Physical laws are usually local, causal influence is strongly correlated with spatial distance, biasing further towards local decentralized, self-organizing computation. Additionally evolution as a soup of mutation and recombination, biases towards finding self-organizing systems that aggregate earlier parts and their functionality, rather than finding a system that rebuilds its components outside-in.

For consciousness to persist over a changing substrate, it needs to be a self-maintaining attractor in brain state space.

Michael Levin’s work suggests an adjacent but more radical idea on consciousness [1, 2]: Consciousness may extend beyond brains, it could be the mechanism that organizes multicellular life on different levels of organization. The same principle that creates and sustains the physical coherence of the organism may also sustain its mental coherence.

Why learning algorithm?

When unconscious, we become vegetative, simple routines like breathing and blood flow continue while planning and more complex actions become impossible. Consciousness seems to be there from the start, it is not something discovered late in child development. During sleepwalking, people usually do not report conscious experience, while being able to walk and interact in simple, yet restricted ways. Memory formation seems to require conscious experience too. You can not remember what happened during anesthesia or deep sleep.

The more routine something becomes, the more it gets automated and the less conscious attention it naturally receives. Changing gears in a car happens without conscious effort. When something unexpected happens, like the gear not locking, the activity becomes prominent again. We can deliberately direct attention, and novelty or surprise captures it. The more surprisal a task contains, the more conscious effort it requires.

Learning can be explained in terms of prediction and compression (see my compression post). Our phenomenal experience is predictive, it is not the representing current case. Experiments like the rubber hand illusion and optical/auditory illusions show this in many ways. Think of a Necker cube alternating between depth interpretations, or mistaking an object in peripheral vision for a person.

Prediction requires understanding of the environment and one’s own system. For that it has to continuously integrate information from sensory input and internal states into its current world model to create a coherent view of what is the case. Competing interpretations in perception have to be resolved, potentially by reciprocal inhibition.

I really like the analogy of consciousness as a consensus algorithm for the mental world, creating a robust identity without centralization, like proof-of-work is for blockchains. Our conscious experience is sequential, a stream of now, experientially we can not experience branching states in the same moment. We experience a story of the last seconds, a coherent frame our consciousness constructed.

The fact one moment of experience binds together multiple features at once is part of binding problem. It is often used as an argument against computationalism, supporting quantum consciousness ideas (see this post).

If we believe consciousness has a causal role, its role is probably continual learning or, as Richard Sutton would point out, what sane people just call learning.

Testing the hypothesis

To test this hypothesis by construction we need a digital self-organizing architecture. Alongside Neural ODEs, reaction-diffusion systems and Ising-inspired models, a promising architecture is the Neural Cellular Automaton (NCA). A Cellular Automaton is mathematically just a topologically continuous shift-equivariant map.

An NCA is a system that learns local rules in training. The learned rule is applied to each node/cell in a graph. The limitation is that current training via BPTT is not self-organizing at all, it is central and offline, only inference is self-organizing. Furthermore, each cell applies exactly the same model/rule, differentation can only happen through heterogeneous initial states and stochastic updating. Practically, differentiating self-organizing systems scale very poorly on current hardware, given we build outside-in hardware into an inside-out world, to run inside-out programs in the outside-in architecture.

Most of those limitations can be alleviated with architectural changes, new ways of training, and simply better training goals. NCA have found many applications in image and texture generation, but can also be used for solving sudokus or implementing distributed matrix multiplication. For them to be applicable to consciousness research, there is still a long way to go.

Consciousness was found by evolution so maybe it is simpler than we assume, or maybe exactly due to its evolutionary origin, it is harder to find via current machine learning methods.

As Chris Olah says, neural networks are grown not built. So the vision is not to build consciousness, it is to grow it through learning on a task that requires continual in-context learning for success (e.g. video prediction). If we train on such a task and recognize a sudden phase transition in its behavior and abilities, recognizing more and more properties related to consciousness (attention-control, self-modeling, reflection…), we might be able to isolate consciousness as a self-organizing learning algorithm, and find out what makes it crystallize.

Credits

I wrote this post in preparation for my talk at the CIMC x Berlin Workshop. Joscha Bach shaped a lot of my thinking in Consciousness Research and Philosophy of Mind, therefore a lot of the ideas and framings in here mix my thoughts with my interpretation of his. I credited any of the explicit ideas directly in this post.