### What is a theory anyway? Let's say you have a theory of motion. It essentially predicts trajectories of balls thrown in the air based on forces that are applied to them. Testing such a theory is simple: you setup an experiment with a ball, apply some forces and then check if the trajectory of motion predicted by your theory matches to actual observes trajectory. If prediction matches observations, you increase your confidence in the theory. If it *drastically* mismatches, you reject your theory. #### But you can always modify your theory to match predictions You can make your theory consistent with observations by adding mismatches in the theory itself (in some complicated fashion). This is what [epicycles theory]([https://www.spitzinc.com › blog › epicycles-and-discovering-bad-theories](https://www.spitzinc.com/blog/epicycles-and-discovering-bad-theories/)) did whenever a mismatch was found between its predictions and observations. In the extreme case, your theory can be simply just a bunch of prediction errors. #### So, how do new theories replace old ones? If you can always modify an old theory with arbitrary tweaks, what causes a new theory to grow in adoption? I think ultimately it boils down to **usefulness**. You adopt a theory that enables you to achieve your goals. If your goals involve predictions in previously unseen situations, that becomes your guiding star. This is the same reason we prefer parsimonious theories. **We like simpler theories because they consume less cognitive resources for the same predictive accuracy.** My claim is that theories take the shape of tools. **They don't reveal reality "as it is" but rather helping us achieve our goals with minimal effort.** ### How consciousness theories are different Let's say you have a theory of consciousness (like integrated information theory). How you test this theory? Perhaps you'd do an experiment like the following one: > Take a physical system X and get a prediction from the theory about what type of consciousness this system has (in terms of "*what is it like to be X*""). Now, verify that the system X is really having an experience like what the theory has predicted. The key question becomes: how would you verify what experience system X is really having? - We can't look at the system like we look at the ball's trajectory and deduce what it's experiencing (when we open up our brain, all we see is neurons and not the beauty of the sunset) - We can't ask the system since it may not use language or be able to report experience or have any motivation to reveal to us what it is experiencing This difficulty in verifying a theory of consciousness on an arbitrary physical system leads to people relying on their intuition about what systems are capable of experiencing a world. For example, regarding integrated information theory (IIT), Scott Aaronson famously critiques that an IIT says a 2D grid of XOR gates is more conscious than humans. ![[Screenshot 2022-01-17 at 12.51.24 PM.png]] His whole talk is worth listening to: <iframe width="560" height="315" src="https://www.youtube.com/embed/R2yRxZCPkws" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> The question of whether an XOR grid has more consciousness than humans is fundamentally unresolvable partly because we can't ask the XOR grid what it is experiencing (and partly we don't understand what does "more consciousness" even mean). #### But we have theories of black holes Physics is no stranger to inaccessible objects. A theory's prediction that cannot be verified doesn't invalidate the theory. We can verify other predictions of theories and _that_ increases our confidence in the theory. For example, even if we can't go in the black hole, we can predict how light will bend around it and observe that light from earth. This gives a test for the theory. Of course, observing the black hole directly will be great but since it's a theoretically inaccessible object, verifying multiple other predictions from the theory increases our confidence about theory's predictions about the black hole (even if we can never directly observe it ourselves). As an aside, this to me suggests a bayesian way of betting on theories: **each additional confirmation should increase confidence in all future predictions of the theory (even if many of such experiments can never be performed).** #### Impossibility of verifying consciousness Just like a theory of black holes makes predictions about how the light bends around them which we can verify, a theory of consciousness can make predictions about how consciousness can impact behavior of a system which we can verify. But even if our theory of consciousness perfectly predicts the behavior, we will always have debates about whether our theory is just modelling behavior instead of the subjective experience. Actually, the issue is similar for black holes. We assumes black holes exist if assuming that black holes exist helps us explain in a simple way all other complicated phenomena such as light bending, galaxy rotation speed, etc. Remember: we prefer simpler models that make high accuracy predictions. Similarly, **if the assumption of a system's agency + inner world helps us explain the behaviour of the system, we should increase our confidence in the belief that the system definitely has agency and an experienced inner world.** Although there will always be critics who'd deny existence of inner world, but as detailed in [[Understanding as hierarchical model building]], all we have access to are models of the reality (and never reality itself). We have to assume Stephen Hawking's [model-dependent realism]([https://en.wikipedia.org › wiki › Model-dependent_realism](https://en.wikipedia.org/wiki/Model-dependent_realism)) because models are all what we have access to. I think we have critics of theories of consciousness because **we don't yet have a theory of consciousness that models a system's experienced world (including its self) to make good predictions about system's behavior.** ### Behavior as a low-dimensional function of experience Through introspection we can see that _all_ the richness of our experienced world seldom translates into behavior/action. We could be laying still in the garden and yet see birds, hear noises, smell fragrances, feel hungry. A rich inner world is lit up but it isn't translating into behavior. This presents a big difficulty in theories of consciousness. We simply can't observe the rich experienced world of naturally occurring systems (except humans) and without that, we would have great difficulty in hypothesizing what aspects of a physical system are essential for consciousness and what aren't. For example, if we are limited to observing humans and how they experience the world, how would we know if neurons are _essential_ part of our theory of consciousness or if they're simply a particular implementation of a much more general thing (which could exist in all sorts of non-biological systems). ### Building a theory of consciousness for humans With humans, we have the benefit of having access to both: behavior and reports of subjective experience. So, perhaps, any theory of consciousness should first be tested against reported experiences and exhibited behavior of humans before extrapolation to non-human systems. If the theory doesn't even work for humans, the only systems that we intuitively trust as having subjective experience that they can communicate, what's the point of applying it to non-humans? One way to resolve the question of what aspects of humans are essential to consciousness is to consider that a theory that models reality well will have more predictions, less error per prediction than a theory that doesn't model reality well. So, if we have two theories: one in which neurons play a central role; another in which neurons are simply a particular instantiation of a more gneral thing discover by evolution, we can pit both of these theories against each other to see which one better predicts experiences and behavior of the humans. Better here means: more aspects of behavior and experiences with less error. Once we have a high confidence in a theory of consciousness as applied on humans, we can apply the same theory on different physical systems and see how theory-predicted behavior matches with observed behavior. Of course, like we don't have access to black holes, we will never have access to inner experiences of the physical system. But **if the theory correctly predicts behaviors of multiple disparate systems and predicts experiences + behavior of humans, we can increase our belief in whatever theory says about inner experience of physical system as actually happening.** #### Challenges with building a theory of consciousness for humans Brain is where consciousness takes place. We know that because lesions in different parts of brains lead to [disorders of various kinds](https://en.wikipedia.org/wiki/The_Man_Who_Mistook_His_Wife_for_a_Hat). We also know that brain is _both_ complicated and complex. There are 86 billion neurons in the adult human brain which collectively have trillions of synapses. We have beautiful, simple theories in physics because the systems we are dealing with can be described using few variables at most. [Black hole has mass, charge and angular momentum](https://en.wikipedia.org/wiki/No-hair_theorem). That's it! You can describe any black hole using these three variables. [[Why fundamental reality is describable using mathematics|We can mathematically model black holes *because* they're so simple.]] Contrast this with the brain with its 86 billion neurons and our rich subjective experience containing millions of qualia datapoints. **When both input and output is so rich, it's evident that we shouldn't expect a simple theory of consciousness that directly maps brain state to experiences.** As I wrote in [[Why we dont have a science of consciousness yet]], one difficulty is that we have extremely sparse data on human brain activity (primarily due to [[Why brain-machine interfaces progress so slowly|ethical reasons]]). But even if we had all the data (brain activity and reported subjective experiences), an elegant theory wouldn't automatically pop out of it. #### Building predictors of human subjective experience So, it seems that a comprehensive theory of consciousness cannot be as compact as, say, [[Demystifying general relativity|general relativity]] because _both_ input variables (brain state) and output predictions (exact subjective experience) are complex and multi-dimensional. Neural networks are famous for mapping these multi-dimensional inputs and output. So given enough data, sure, we would be able to build _predictors_ that map brain activity to subjective experiences and those would be a step in the right direction. But such predictors will a) be likely as mysterious and "black box" as the brain itself; b) would apriori assume neurons and their spiking activity as the only substrate capable of generating subjective experiences. This would preclude application of these predictors to aneural systems (such as silicon based hardware, algorithms and non-neural biological systems). ### Types of theories of consciousness So, it seems we are at a crossroads where two types of theories of consciousness are possible. 1. Either we have predictions based on neural correlates of consciousness for humans (which we likely won't be able to confidently apply to aneural systems and which likely won't be particularly satisfying to our understanding) 2. Or, we have mathematically elegant and simple theories like integrated information theory that ultimately you can't verify precisely and hence remain always open to debate because: - Their outputs are limited in their richness while our conscious world is rich (e.g. IIT primarily predicts level of consciousness as phi) - Barring a comprehensive prediction of our subjective experience, our confidence in theory remains low and hence we're hesitant of applying the theory widely So, arrive at the upshot of the essay: > As far as theories of consciousness go, either we will have highly accurate predictive models as complex as the brain itself that we can't easily understand OR we will have compact mathematical models that don't predict that much about our conscious experience (and hence remain doubtful/unverified) Physics theories work because we're able to apply them to lots of different systems (balls, planets, black holes) and verify that such theories work generally. Imagine if a theory of gravitation _only_ worked for calculating the attraction between the moon and the earth. Having such a theory would be better than not having any theory but certainly it wouldn't compete with general relativity which can be applied widely. ### The difficult path of building theories of consciousness So, theories of consciousness face similar problem: - We have to take humans as initial subjects of study because they can report their conscious states - We don't have enough data on human brain activity due to ethical concerns - Even when we have enough data, simpler models of the data will not generate enough precise predictions and hence remain debatable as saying anything general about consciousness (as we say Scott deflating IIT) and successful predictors of experiences from brain activity will be black box and preclude aneural systems - Lacking a generally accepted theory of consciousness, we will always have a debate about which physical systems are conscious Another problem with porting of theories of consciousness, **a physical system's experienced world could be completely different than ours** (on which theories of consciousness is built). Love, emotion, sound, etc. is all what we humans are aware of. What is a system like an "ant colony aware of"? What is the equivalent of awake for a single celled organism? ### Summing up For a theory of consciousness that takes some inputs (such as brain states, or information network) and outputs something about subjective experience, we will have great difficulty in knowing whether we've captured something universal OR whether we've simply captured a correlate/instantiation of such universal feature in humans. - Correlates are NOT exciting because they're not generalizable - Generealizable are NOT verifiable because humans are the only ones who can report experiences In sum, our confidence in a theory of consciousness will never be at part with general relativity or quantum mechanics that are universally applicable and verifiable from many different sources and have a simple input -> output map that we can claim to understand at some level. ### Sidebar: what about the hard problem of consciousness? All what we've talked about so far is the ["real" problem of consciousness](https://aeon.co/essays/the-hard-problem-of-consciousness-is-a-distraction-from-the-real-one). That is, given an input state, how can we predict an output state that is experienced world. There's also the "hard" problem of consciousness that asks how do mental qualities / qualia (the redness of red, sweet smell of vanilla, etc.) actually get generated by non-mental physical material (say neurons)? I think this problem is either confused or impossible to answer because **qualia are the atoms of understanding** [[Understanding as hierarchical model building|on top of which all other understanding of the world is built]]. For example, when we learn to understand 1+1 = 2, we're visualising how it feels for two individual objects (such as apples) to come together in our experienced world (via any of the senses such as visual, touch, etc.) Qualia is the building block of our reality and hence we can't understand it in any other term. In fact, **we understand everything else in terms of qualia.** ### So, what to do? I think since our hopes for fully general theory of consciousness that can be universally applied is limited, we should focus on approaching theory-building from usefulness perspective. Instead of asking what theory is true (because it's inaccessible), we should ask: **what will be a useful theory of consciousness?** The fact that theories in physics also broadly track usefulness (which is measured by accuracy), leaning towards usefulness as a heuristic will help us make concrete progress and actually solve problems in the world. Although the lack of a universally accepted theory of consciousness will legitmize violence against [[Sentient beings that are capable of suffering]], **perhaps our progress in the science of consciousness will put pressure on people to broaden their moral horizon beyond humans** (at least, that is one of the strong motivators to push the theory of consciousness). <iframe class="signup-iframe" src="https://invertedpassion.com/signup-collector" title="Signup collector"></iframe>