Snippets from Mike Johnson’s Principia Qualia (2016) for my own future reference. The Principia Qualia “attempts to chart a viable course toward a general theory of valence, a.k.a. universal, substrate-independent principles that apply equally to and are precisely true in all conscious entities, be they humans, non-human animals, aliens, or conscious AI”.
Johnson claims that while valence in the human brain is a complex phenomenon which defies simple description (as summarised in Fig 1 below), valence itself isn’t necessarily a complex phenomenon; instead he thinks it’s “a crisp thing we can quantify, and the patterns in it only look incredibly messy because we’re looking at it from the wrong level of abstraction”.
The right level of abstraction involves “understanding valence research as a subset of qualia research, and qualia research as a problem in information theory and/or physics, rather than neuroscience”, hence entails seeking “general principles for how the causal organization of a physical system generates or corresponds to its phenomenology, or how it feels to subjectively be that system”. This in turn requires addressing more general questions of consciousness and qualia. Johnson rejects top-down theories of consciousness (constructed around how it feels and the high-level dynamics of how the brain implements what we experience) because (1) “we have no reason to expect that our high-level internal phenomenology has any crisp, intuitive correspondence with the underlying physics and organizational principles which give rise to it”, which suggests “theories of consciousness or valence which take high-level psychological concepts as primitives will be “leaky abstractions””, and (2) they’re not always testable. The set of bottom-up theories of consciousness (Johnson’s rule of thumb: “it should apply clearly and equally to any arbitrary cubic foot of space-time, and offer testable predictions at multiple levels of abstraction”) for now has only one subset, Tononi’s IIT and IIT-inspired approaches (Tegmark’s Perceptronium approach and Barrett’s FIIH), so he spends a good portion of the paper analysing them.
He ultimately concluded that IIT is an “extremely promising approach for deriving the “data structure isomorphic to what it feels like to be a system” – but it’s also deeply flawed or underdeveloped in certain details”. (As a blast to the past, Johnson references Scott Aaronson’s Why I Am Not An Integrated Information Theorist (or, The Unconscious Expander) and the discussion below, which I remember reading back in university and considering it sufficiently persuasive to write off IIT.) One property he likes about IIT, which he preserves in his hypothesis below, is that it aims at PHP4, which IIT formalises by aiming to construct a mathematical object isomorphic to a system’s qualia; it’s “a clear and simple goal, and a great antidote to much of the murky confusion surrounding consciousness research (footnote: if this is the right goal, then we can get on with trying to achieve it. If it’s the wrong goal, the onus is now on critics to explain why)”.
Since valence is a subset of phenomenology, “reverse-engineering valence is simply a matter of figuring out how valence is encoded within the mathematical object constructed by IIT to be isomorphic to the system’s phenomenology”.
This quote was pretty funny:
The best proximate solution for improving IIT would involve locking Tononi, Koch, Aaronson, Griffith, Tegmark, and maybe David Spivak (a leading expert on Grothendieck topology) in a room, and not letting them out until everybody’s satisfied with the math. But I actually think there’s another option that’s much simpler, more effective in the long-term, and less likely to lead to police reports.
Below is the alternative option. Johnson claims that a full-stack quantitative solution to consciousness, as distinct from the narrower research programs currently active in the field (Tononi’s IIT, Tsuchiya’s lab, Tegmark’s Perceptronium approach / Barrett’s FIIH), has to involve “three distinct core tasks”:
Unpacking and mapping these sub-problems yields the following:
How popular consciousness theories fare when evaluated against the rubric defined by the sub-problems above:
An aside on why it’s proven very hard to make progress on the substrate problem (“Which subset of objects & processes in our chosen ontology [in this case, physics] ‘count’ toward consciousness?”):
Three principles Johnson believes is necessary for a satisfying mathematical derivation of valence, drawing from the framework above (with the first two principles being necessary for any satisfying solution to consciousness):
An aside on qualia, and why self-reports can’t be the gold-standard ground-truth about them:
What assuming these three principles buys us:
Heuristics for zeroing in on the precise mathematical property that corresponds to valence:
Slightly less than two degrees of freedom seem necessary & sufficient to describe valence space:
A bit more on the translation problem:
Given a mathematical object isomorphic to a system’s phenomenology, how do we populate a translation list between its mathematical properties and the part of phenomenology each property or pattern corresponds to?
From Johnson’s suggestions above (that qualia may be simple vs complex, atomic vs composite, and local vs global, intuitively important vs intuitively trivial, and these distinctions will apply equally to both qualia and their corresponding geometric representation) he infers that “knowledge about qualia will (at least initially) take the form of a translation table between geometric/topological properties, and their corresponding phenomena in the qualia domain”, and then tabulates these “very speculative possible starting points, of unknown quality… described at a high and lossy level of abstraction”:
Johnson points out an interesting implication of his hypothesis applied to IIT-like consciousness theories: maximising pleasure/symmetry makes consciousness/Φ drop very rapidly, “i.e. a fully symmetric system has no room for the sort of complexity necessary for integrated information. This is consistent with the phenomenological observation that pleasurable experiences/flow states involve time compression, and vice-versa”:
This provides some guidance on how to define the more general state-space of valence × intensity × consciousness, as per Fig 13 below. A system’s morally significant hedonic status (its ‘happiness’ or ‘suffering’) is then its location within the state-space below:
On how Johnson’s hypothesis above pays rent, Section XI discusses 3 experimental tests doable “now”, and Section XII ventures some downstream hypotheses:
On the anatomy & network topology of valence:
“hedonic’ brain regions influence mood by virtue of acting as ’tuning knobs’ for symmetry/harmony in the brain’s consciousness centers. Likewise, nociceptors, and the brain regions which gate & interpret their signals, will be located at critical points in brain networks, able to cause large amounts of salience-inducing antisymmetry very efficiently”
“We should also expect rhythm to be a powerful tool for modeling brain dynamics involving valence- for instance, we should be able to extend (Safron 2016)’s model of rhythmic entrainment in orgasm to other sorts of pleasure. Footnote: My expectation is that people with musical anhedonia (∼2-5% of the population) may have an abnormal auditory preprocessing pipeline that significantly alters the structure of auditory stimuli before it gets to a consciousness center. I.e., music effectively ’hacks’ valence partly because it’s highly patterned (symmetrical), and people who have musical anhedonia will have a non-standard ear-to-thalamus pipeline that breaks these sorts of patterns.”
On valence & neuropharmacology:
“Non-opioid painkillers and anti-depressants are complex, but it may turn out that a core mechanism by which they act is by introducing noise into neural activity and connectivity, respectively. This would explain the odd findings that acetaminophen blunts acute pleasure (Durso, Luttrell, and Way 2015), and that anti-depressants can induce long-term affective flattening. Footnote: My friend Andres Gomez Emilsson suggests this could be due to upregulation of agmatine (personal discussion)”
“Psychedelic substances, although often pleasurable, actually increase emotional variance by biasing the brain toward symmetrical structure, and could result in enhanced pain if this structure is then broken – i.e., they are in this sense the opposite of painkillers”
“Severe tinnitus could lead to affective flattening for similar interference-based reasons: insofar as the brain’s subconscious preprocessing can’t tune it out, the presence of a constant pattern in consciousness would likely make it more difficult to generate symmetries (valence) on-the-fly. This would also imply that the specific frequency pattern of the perceived tinnitus sensation may matter more than is commonly assumed”
On self-organization & deep learning, and how we aren’t pleasure-maximisers strictly speaking:
“My hypothesis implies that symmetry/harmony is a core component of the brain’s organizational & computational syntax: specifically, we should think of symmetry as one (of many) dynamic attractors in the brain. Footnote: By implication, humans are not strictly speaking pleasure-maximizers… but we do tend to work to increase our valence, satisficing against our other dynamic attractors. This suggests that mammals got a bit lucky that we evolved to seek out pleasure! But not that lucky, since symmetry is a very functionally-relevant and useful property for systems to self-organize around, for at least two reasons:
“First, self-organizing systems such as the brain must develop some way to perform error-correction, measure & maintain homeostasis, and guide & constrain morphological development. Symmetry-as-a-dynamic-attractor is a profoundly powerful solution to all of these which could evolve in incrementally-useful forms, and so symmetry-seeking seems like a common, perhaps nigh-universal evolutionary path to take. Indeed, it might be exceedingly difficult to develop a system with complex adaptive traits without heavy reliance upon principles of symmetry”
“Second, the brain embodies principles of symmetry because it’s an efficient structure for modeling our world. (Lin and Tegmark 2016) note that physics and deep learning neural networks display cross-domain parallels such as “symmetry, locality, compositionality and polynomial log-probability,” and that deep learning can often avoid combinatorial explosion due to the fact that the physical world has lots of predictable symmetries, which enable unusually efficient neural network encoding schemes”
On boredom as a sophisticated “anti-wireheading” technology:
“Why do we find pure order & symmetry boring, and not particularly beautiful? I posit boredom is a very sophisticated “anti-wireheading” technology which prevents the symmetry/pleasure attractor basin from being too ‘sticky’, and may be activated by an especially low rate of Reward Prediction Errors (RPEs). Musical features which add mathematical variations or imperfections to the structure of music – e.g., syncopated rhythms (Witek et al. 2014), vocal burrs, etc – seem to make music more addictive and allows us to find long-term pleasure in listening to it, by hacking the mechanic(s) by which the brain implements boredom”
As far as I can tell, no, he instead means how things feel, e.g. the pleasure / pain / neutral triangle continuum above. Here he summarises his Principia Qualia like so:
PQ begins by considering a rather modest question: what is emotional valence? What makes some things feel better than others?
This sounds like the sort of clear-cut puzzle affective neuroscience should be able to solve, yet all existing answers to this question are incoherent or circular. Giulio Tononi’s Integrated Information Theory (IIT) is an example of the kind of quantitative theory which could in theory address valence in a principled way, but unfortunately the current version of IIT is both flawed and incomplete. I offer a framework to resolve generalize IIT by distilling the problem of consciousness into eight discrete & modular sub-problems (of which IIT directly addresses five).
Finally, I return to valence, and offer a crisp, falsifiable hypothesis as to what it is in terms of something like IIT’s output, and discuss novel implications for neuroscience.
Snippets from Mike Johnson’s Principia Qualia (2016) for my own future reference. The Principia Qualia “attempts to chart a viable course toward a general theory of valence, a.k.a. universal, substrate-independent principles that apply equally to and are precisely true in all conscious entities, be they humans, non-human animals, aliens, or conscious AI”.
Johnson claims that while valence in the human brain is a complex phenomenon which defies simple description (as summarised in Fig 1 below), valence itself isn’t necessarily a complex phenomenon; instead he thinks it’s “a crisp thing we can quantify, and the patterns in it only look incredibly messy because we’re looking at it from the wrong level of abstraction”.
The right level of abstraction involves “understanding valence research as a subset of qualia research, and qualia research as a problem in information theory and/or physics, rather than neuroscience”, hence entails seeking “general principles for how the causal organization of a physical system generates or corresponds to its phenomenology, or how it feels to subjectively be that system”. This in turn requires addressing more general questions of consciousness and qualia. Johnson rejects top-down theories of consciousness (constructed around how it feels and the high-level dynamics of how the brain implements what we experience) because (1) “we have no reason to expect that our high-level internal phenomenology has any crisp, intuitive correspondence with the underlying physics and organizational principles which give rise to it”, which suggests “theories of consciousness or valence which take high-level psychological concepts as primitives will be “leaky abstractions””, and (2) they’re not always testable. The set of bottom-up theories of consciousness (Johnson’s rule of thumb: “it should apply clearly and equally to any arbitrary cubic foot of space-time, and offer testable predictions at multiple levels of abstraction”) for now has only one subset, Tononi’s IIT and IIT-inspired approaches (Tegmark’s Perceptronium approach and Barrett’s FIIH), so he spends a good portion of the paper analysing them.
He ultimately concluded that IIT is an “extremely promising approach for deriving the “data structure isomorphic to what it feels like to be a system” – but it’s also deeply flawed or underdeveloped in certain details”. (As a blast to the past, Johnson references Scott Aaronson’s Why I Am Not An Integrated Information Theorist (or, The Unconscious Expander) and the discussion below, which I remember reading back in university and considering it sufficiently persuasive to write off IIT.) One property he likes about IIT, which he preserves in his hypothesis below, is that it aims at PHP4, which IIT formalises by aiming to construct a mathematical object isomorphic to a system’s qualia; it’s “a clear and simple goal, and a great antidote to much of the murky confusion surrounding consciousness research (footnote: if this is the right goal, then we can get on with trying to achieve it. If it’s the wrong goal, the onus is now on critics to explain why)”.
Since valence is a subset of phenomenology, “reverse-engineering valence is simply a matter of figuring out how valence is encoded within the mathematical object constructed by IIT to be isomorphic to the system’s phenomenology”.
This quote was pretty funny:
Below is the alternative option. Johnson claims that a full-stack quantitative solution to consciousness, as distinct from the narrower research programs currently active in the field (Tononi’s IIT, Tsuchiya’s lab, Tegmark’s Perceptronium approach / Barrett’s FIIH), has to involve “three distinct core tasks”:
Unpacking and mapping these sub-problems yields the following:
How popular consciousness theories fare when evaluated against the rubric defined by the sub-problems above:
An aside on why it’s proven very hard to make progress on the substrate problem (“Which subset of objects & processes in our chosen ontology [in this case, physics] ‘count’ toward consciousness?”):
Three principles Johnson believes is necessary for a satisfying mathematical derivation of valence, drawing from the framework above (with the first two principles being necessary for any satisfying solution to consciousness):
An aside on qualia, and why self-reports can’t be the gold-standard ground-truth about them:
What assuming these three principles buys us:
Heuristics for zeroing in on the precise mathematical property that corresponds to valence:
Slightly less than two degrees of freedom seem necessary & sufficient to describe valence space:
A bit more on the translation problem:
From Johnson’s suggestions above (that qualia may be simple vs complex, atomic vs composite, and local vs global, intuitively important vs intuitively trivial, and these distinctions will apply equally to both qualia and their corresponding geometric representation) he infers that “knowledge about qualia will (at least initially) take the form of a translation table between geometric/topological properties, and their corresponding phenomena in the qualia domain”, and then tabulates these “very speculative possible starting points, of unknown quality… described at a high and lossy level of abstraction”:
Johnson points out an interesting implication of his hypothesis applied to IIT-like consciousness theories: maximising pleasure/symmetry makes consciousness/Φ drop very rapidly, “i.e. a fully symmetric system has no room for the sort of complexity necessary for integrated information. This is consistent with the phenomenological observation that pleasurable experiences/flow states involve time compression, and vice-versa”:
This provides some guidance on how to define the more general state-space of valence × intensity × consciousness, as per Fig 13 below. A system’s morally significant hedonic status (its ‘happiness’ or ‘suffering’) is then its location within the state-space below:
On how Johnson’s hypothesis above pays rent, Section XI discusses 3 experimental tests doable “now”, and Section XII ventures some downstream hypotheses:
On the anatomy & network topology of valence:
“hedonic’ brain regions influence mood by virtue of acting as ’tuning knobs’ for symmetry/harmony in the brain’s consciousness centers. Likewise, nociceptors, and the brain regions which gate & interpret their signals, will be located at critical points in brain networks, able to cause large amounts of salience-inducing antisymmetry very efficiently”
“We should also expect rhythm to be a powerful tool for modeling brain dynamics involving valence- for instance, we should be able to extend (Safron 2016)’s model of rhythmic entrainment in orgasm to other sorts of pleasure. Footnote: My expectation is that people with musical anhedonia (∼2-5% of the population) may have an abnormal auditory preprocessing pipeline that significantly alters the structure of auditory stimuli before it gets to a consciousness center. I.e., music effectively ’hacks’ valence partly because it’s highly patterned (symmetrical), and people who have musical anhedonia will have a non-standard ear-to-thalamus pipeline that breaks these sorts of patterns.”
On valence & neuropharmacology:
“Non-opioid painkillers and anti-depressants are complex, but it may turn out that a core mechanism by which they act is by introducing noise into neural activity and connectivity, respectively. This would explain the odd findings that acetaminophen blunts acute pleasure (Durso, Luttrell, and Way 2015), and that anti-depressants can induce long-term affective flattening. Footnote: My friend Andres Gomez Emilsson suggests this could be due to upregulation of agmatine (personal discussion)”
“Psychedelic substances, although often pleasurable, actually increase emotional variance by biasing the brain toward symmetrical structure, and could result in enhanced pain if this structure is then broken – i.e., they are in this sense the opposite of painkillers”
“Severe tinnitus could lead to affective flattening for similar interference-based reasons: insofar as the brain’s subconscious preprocessing can’t tune it out, the presence of a constant pattern in consciousness would likely make it more difficult to generate symmetries (valence) on-the-fly. This would also imply that the specific frequency pattern of the perceived tinnitus sensation may matter more than is commonly assumed”
On self-organization & deep learning, and how we aren’t pleasure-maximisers strictly speaking:
“My hypothesis implies that symmetry/harmony is a core component of the brain’s organizational & computational syntax: specifically, we should think of symmetry as one (of many) dynamic attractors in the brain. Footnote: By implication, humans are not strictly speaking pleasure-maximizers… but we do tend to work to increase our valence, satisficing against our other dynamic attractors. This suggests that mammals got a bit lucky that we evolved to seek out pleasure! But not that lucky, since symmetry is a very functionally-relevant and useful property for systems to self-organize around, for at least two reasons:
“First, self-organizing systems such as the brain must develop some way to perform error-correction, measure & maintain homeostasis, and guide & constrain morphological development. Symmetry-as-a-dynamic-attractor is a profoundly powerful solution to all of these which could evolve in incrementally-useful forms, and so symmetry-seeking seems like a common, perhaps nigh-universal evolutionary path to take. Indeed, it might be exceedingly difficult to develop a system with complex adaptive traits without heavy reliance upon principles of symmetry”
“Second, the brain embodies principles of symmetry because it’s an efficient structure for modeling our world. (Lin and Tegmark 2016) note that physics and deep learning neural networks display cross-domain parallels such as “symmetry, locality, compositionality and polynomial log-probability,” and that deep learning can often avoid combinatorial explosion due to the fact that the physical world has lots of predictable symmetries, which enable unusually efficient neural network encoding schemes”
On boredom as a sophisticated “anti-wireheading” technology:
“Why do we find pure order & symmetry boring, and not particularly beautiful? I posit boredom is a very sophisticated “anti-wireheading” technology which prevents the symmetry/pleasure attractor basin from being too ‘sticky’, and may be activated by an especially low rate of Reward Prediction Errors (RPEs). Musical features which add mathematical variations or imperfections to the structure of music – e.g., syncopated rhythms (Witek et al. 2014), vocal burrs, etc – seem to make music more addictive and allows us to find long-term pleasure in listening to it, by hacking the mechanic(s) by which the brain implements boredom”
Some of Johnson’s older (2015) essays:
Effective Altruism, and building a better QALY
How understanding valence could help make future AIs safer
Sounds interesting. What does he mean by valence? Is it in the neighborhood of wanting or valuing?
As far as I can tell, no, he instead means how things feel, e.g. the pleasure / pain / neutral triangle continuum above. Here he summarises his Principia Qualia like so: