PrincipiaQualia

Principia Qualia: Blueprint for a new science (usually just "PrincipiaQualia") is a short book by Michael Edward Johnson written in 2016. It outlines a full-stack research paradigm for consciousness and inspired the founding of QRI two years later.

The book is structured into three main sections ("Review", "Valence", "Discussion") plus a list of appendices. The overarching focus is the role and nature of Valence. This article will take a liberal approach in summarizing the book, with uneven space given to different sections. The text will frequently provide page numbers so readers can consult the book for further details, which is freely accessible on Mike's blog OpenTheory.

In 2023 (after his departure from QRI), Mike published a follow-up work titled "Qualia Formalism and a Symmetry Theory of Valence". This paper will not be covered separately since it is primarily an abbreviated version of PrincipiaQualia but will sporadically be referenced. Note that many sections that are only briefly covered here are omitted in the follow-up paper as well.

A recurring theme throughout the book is to take philosophical assumptions seriously (such as Consciousness Realism and Qualia Structuralism) and carefully think through their implications for a proper scientific theory of consciousness. A sizeable chunk of the arguments made throughout the book are derived from this approach, though it also engages extensively with the academic literature.

Review (Book Section)

The Neuroscience Component

The book opens with a summary of the work on valence in the literature (pp 1-5). One prevalent school of thought aims to understand valence by analyzing its functional behavior, e.g., as a reinforcement learning signal, or as a way to regulate the brain's exploration and risk-taking. For example, being happy tends to make one more open-minded to novel experiences, whereas a depressed mood tends to cause a defensive attitude focused on avoiding danger. In this way, assessing valence allows the brain to choose a strategy appropriate for the given context.

However, functional stories about valence tend to have exceptions, and thus, be heuristics rather than brute facts. Furthermore, there is evidence that wanting and liking are separate processes in the brain (p. 2), and there are cases where they come apart. Examples of both extremes are drug addiction on the one hand (high wanting but low liking) and advanced meditative states on the other (high liking but low wanting). Mike summarizes the situation as follows (p. 1, emphasis in all quotes from the original):

Affective neuroscience has been very effective at illuminating the dynamics and correlations of how valence works in the human brain, on a practical level, and what valence is not, on a metaphysical level. This is useful yet not philosophically rigorous, and this trend is likely to continue.

Another class of results relates to the localization of pain and pleasure in the brain. In this case, the data suggests that

• the origin of both pain and pleasure can, to an extent, be traced back to certain regions of the brain; where
• pain seems to be the more distributed phenomenon (i.e., with more regions contributing) but also easier to cause; and
• none of these findings are absolute (e.g., the regions responsible for causing pleasure don't always do so; some are involved in both pleasure and pain, and so on).

Mike concludes (p. 3):

Importantly, the key takeaway from the neuro-anatomical research into valence is this: at this time we don’t have a clue as to what properties are necessary or sufficient to make a given brain region a so-called “pleasure center” or “pain center.” Instead, we just know that some regions of the brain appear to contribute much more to valence than others.

Similarly, when distinguishing different types of valence, the taxonomies suggested in the literature seem substantially artificial and metaphysically unsatisfying. One finding Mike singles out (p. 4) is that "[...] pain is easy to cause, but hard to localize in the brain; pleasure has a more definite footprint in the brain, but is much harder to generate on demand".

A summary of the limitations of mechanic understandings of valence from the literature

The following section (pp. 4-5) discusses more technical issues with neuroscience research. These will be skipped here, but see the graphic to the right for a summary. Mike concludes that "our current understanding of valence and consciousness is extremely limited" and suggests that "the core hurdle for affective neuroscience is philosophical confusion, not mere lack of data", thus justifying a more radical rethinking of the problem.

The Philosophical Component

The core premise on which the book's approach is built is to treat valence as a well-defined physical property. This approach relies on a realist view of consciousness, under which valence is a frame-invariant property of all conscious systems. Given this premise, a solution to valence should look more like an equation and less like a bundle of heuristics. Mike also frames this distinction in terms of bottom-up vs. top-down theories and argues that a bottom-up approach is needed.

He singles out Integrated Information Theory (IIT) as the leading bottom-up theory in the literature. This is followed by an extensive discussion of IIT (pp. 8-17) that will be omitted here since it largely aligns with the strengths and problems of IIT discussed in its wiki article. The discussion on IIT-inspired approaches (pp. 17-20) will also be skipped. The upshot in the book is similar to the one drawn by QRI: that the goal of IIT (to construct a mathematical object that precisely describes a moment of consciousness) is desirable even if its precise math is incorrect.

The Subproblems of Consciousness

The decomposition of consciousness into eight subproblems suggested in PrincipiaQualia (p. 23)

Inspired by the approach of IIT, Mike suggests a decomposition of the research problem into eight subproblems (see graphic to the right), which still aligns well with QRI's thinking today. He writes (p. 23):

My claim is that these eight sub-problems are necessary to solve consciousness, and in aggregate, may be sufficient.

And also notes that (p. 24):

[...] we can see why IIT is so impressive: using only 1-2 mechanics it can address all of Step 2’s subproblems. However, we also see why Tegmark is interested in grounding IIT in the ontology of physics: right now IIT ignores nuances of the Reality Mapping Problem, which implies that application of IIT to real physical systems will always have a cloud of ambiguity around it.

Valence (Book Section)

Principles

Mike opens this section by proposing three principles for a "mathematical derivation of valence" (p. 25):

• Qualia Formalism states that "for any given conscious experience, there exists a mathematical object isomorphic to its phenomenology" (thus justifying the goal of IIT).
• Qualia Structuralism states that "this mathematical object has a rich set of formal structures".
• Valence Realism states that "valence is a crisp phenomenon of conscious states upon which we can apply a measure".

Mike briefly considers the inverses of each principle (pp. 26-27) and concludes that they are not plausible. Note that QRI still subscribes to all three principles today.

Having established that we can think of qualia as represented by a mathematical object with significant structure – of which valence is a property – Mike argues that this itself is enough to draw significant conclusions (pp. 27-28):

[...] an isomorphism between a structured (e.g., topological, geometric) space and qualia implies that any clean or useful distinction we can make in one realm automatically applies in the other realm as well. And if we can explore what kinds of distinctions in qualia we can make, we can start to chart the explanation space for valence (what ‘kind’ of answer it will be).

He proposes four distinctions (p. 28) that apply equally in qualia space and its mathematical representation:

1. global vs. local
2. simple vs. complex
3. atomic vs. composite
4. intuitively important vs. intuitively trivial

Applying these four axes to valence, he concludes that valence is global, simple, atomic, and intuitively important, thus providing the primary piece of evidence for the Symmetry Theory of Valence (STV).

The Symmetry Theory of Valence

Main Article: Symmetry Theory of Valence

Mike phrases the STV as follows (p. 35):

Given a mathematical object isomorphic to the qualia of a system, the mathematical property which corresponds to how pleasant it is to be that system is that object’s symmetry.

Thus, the STV identifies well-being, or valence, as a quantitative and frame-invariant feature of consciousness. It underlies much of QRI's thinking, such as the focus on phenomenal character rather than narrative content, and the importance of resonance (see also the article on Brain Eigenmodes).

Evidence for the STV

In addition to the structural argument mentioned above (i.e., the fact that symmetry, like valence, seems global, simple, atomic, and intuitively important), Mike goes through five other clusters of evidence for the STV in the book (pp. 30-35). The following provides a brief summary:

• Empirical hints from affective neuroscience: the STV suggests that high valence states require more coordination than low-valence states, aligning with the empirical finding that causing pleasure seems to require more specialized regions. Furthermore, symmetry seems like a useful property for building brains, corresponding to the fact that we seem drawn to high-valence states.
• A priori hints from phenomenology: the STV makes predictions supported by phenomenology, such as the fact that extreme states (i.e., very high or very low valence) tend to be informationally sparse.
• Hints from neurocomputational syntaxes: while neuroscience is yet to fully uncover the brain's computational architecture and syntax, the evidence for resonance and oscillations in the brain (as well as the Free Energy Principle) align well with the STV.
• The Non-adaptedness principle: since the brain was designed by evolution, any given stimulus ought to be pleasant to the extent that seeking it out is good for inclusive genetic fitness. Therefore, any stimulus that diverges from this prediction (i.e., is either more or less pleasant than evolution would predict) might hint at the physical nature of valence. Confusion and cognitive dissonance are negative examples (i.e., a surprisingly low-valence stimulus), whereas music, abstract mathematics, and flow states are positive examples.
• Common patterns across physical formalisms (lessons from physics): when looking at how physics "represents" quantities, symmetry seems like an a priori plausible candidate for valence. This argument is elaborated on later in the book (pp. 37-38).

Quantification

A graphic illustrating reflection symmetries for different shapes. Note that for every dotted line, reflecting the object along that line leaves the object unchanged. The symmetries are also exhaustive: reflecting any shape along a line not drawn would constitute a change to it.

As a preliminary concept, note that symmetry, in general, is formalized as the property of an object that does not change if a transformation is applied to it. For example, a square is both rotation- and reflection-symmetrical (for specific angles and rotation axes) because applying either transformation to it yields an identical square.

In the updated paper, Mike writes the following on quantifying symmetry:

The size of an object’s symmetry group is the gold standard for symmetry metrics and will be the standard for STV. However, calculating this exactly [...] may be epistemologically and computationally intractable for the foreseeable future: we don’t have examples to work from, and for large objects, the analysis gets very complicated. One notable dynamic here is that as we add additional nodes to an object, the size of the object’s symmetry group tends to plummet rapidly; it only takes one point out of position to break a symmetry. On the other hand, as we add more nodes to an object this tends to increase the underlying dimensionality of the object, and as this dimensionality increases the number of possible symmetries skyrockets because there are radically more available mathematical transformations. It’s unclear which process ‘wins’ this race in higher dimensions, and the dimensionality of human-scale minds is likely to be very high [...]

The symmetry group of an object is the mathematical group of transformations under which the object is invariant. In addition to the size of the symmetry group, Mike suggests cost functions and compressibility (which is a general substitute for symmetry since symmetries inherently come with information-theoretical redundancy).

Testing the Hypothesis

At the end of the chapter, Mike suggests three avenues for testing the STV:

• Measuring the compressibility of brain states (more pleasant states should be more compressible)
• Injecting signals directly into the brain (more symmetrical/harmonious signals should feel much better)
• Using vagus nerve stimulation to approximate this effect more cheaply

The details (pp. 41-46) will be omitted here since they are rather technical.

The State Space of Valence

A graphical illustration of the state space of valence

Qualia seems to exist with high, low, or neutral valence. Under the STV, these correspond to symmetry, antisymmetry, and asymmetry, respectively. Mike lists a major chord, nails on a chalkboard, and white noise as examples for each category.

It's worth noting that quantifying harmony is nontrivial, even for a single musical chord. The folk wisdom that small integer ratios (e.g., 2:3 or 2:5) of frequencies lead to consonant chords is problematic since any simple fraction (e.g., ${\frac {1}{2}}$) has infinitely many arbitrarily complex fractions close to it (e.g., $\textstyle {\frac {124445}{248889}}$). In general, relying on ratios yields a discontinuous metric; however, there has been work on developing continuous metrics, one of which is cited in the book.

Discussion (Book Section)

After a brief summary of other contemporary qualia research (all adjacent to Integrated Information Theory), Mike argues that the STV outperforms rival approaches in terms of internal consistency, consistency with other scientific research, proper motivation (i.e., validity of the principles it's based on), and the empirical predictions it makes. He also briefly sketches the implications of the STV on neuroscience (pp. 48-50).

The final section is devoted to the implications of consciousness research on AI safety. For those who believe that AI will come to dominate the future in whatever fashion, the connection between both fields may be the decisive factor for whether consciousness research is worth caring about. Mike writes (pp. 50-51):

If we define ethics as the study of what is intrinsically valuable, it would be a notable understatement to say that understanding consciousness & valence seems critically important for having a non-insane theory of ethics. The task of understanding the good – and treating sentient creatures better – both seem to require understanding valence [...]

Another particularly pressing problem in ethics is AI safety’s ‘Value Problem’, or the question of what normative values we should instill into future artificial intelligences in order to make them friendly to humans. Currently, the state of the art here is indirect normativity, or building systems to teach artificial intelligences what humans value based on watching how humans behave & interact. Unfortunately, this is prone to problems common to machine learning paradigms (e.g., overfitting, proper model selection, etc) as well as the problem of human values being inherently fuzzy and internally inconsistent.

[...]

Research on consciousness & valence by itself won’t solve issues of ethics, AI safety, personal identity, meaning, social health, and how to use the atoms in our light-cone, but this research does seem centrally necessary for good answers to these questions, and provides a way to cut through confusion and build useful tools. [...]

And if the arguments about consciousness and valence in this work are substantially correct, we should be approaching both ethics and AI safety research very differently than we are now.

Although this isn't spelled out in the book, it's worth mentioning that consciousness research (both from Mike and from QRI today) also has implications for the computational architecture of the brain and, as a consequence, on AI timelines and takeoff scenarios. By and large, the literature on both AI capability and safety tends to view the problem through a computational lens, whereas various results from consciousness research suggest that nontrivial physical mechanisms play a role in the brain. If true, the precise implications for AI significantly depend on whether and how easily general intelligence can be achieved anyway in the current paradigm. (See the Electromagnetic Hypothesis and the article on Computation for more on this point.)

Appendices (Book Section)

The appendices (pp. 57-82) are structured into five sections that each address one of the subproblems of consciousness introduced in the graphic above (missing the Scale-, Boundary- and Topology of Information Problem from the Math section), plus a sixth part speculating on the big picture implications of everything. They won't be covered here since they're not integral to the core argument, but readers are encouraged to consult the book directly.