A significant fraction of the EA movement is concerned with suffering, and all else being equal, think there should be less of it. I think this is an extraordinarily noble goal.
But what *is* suffering? There are roughly as many working definitions of suffering in the EA movement as there are factions in the EA movement. Worryingly, these definitions often implicitly or explicitly conflict, and only the fact that the EA pie is growing relatively rapidly prevents a descent into factional warfare over resources being wasted on ‘incorrect’ understandings of suffering.
Intuitively, one would hope that gradual progress in affective neuroscience will make this problem less pressing- that given enough time&effort&resources, different approaches to defining suffering will cohere, and this problem will fade away.
I am here to inform you that this is not going to happen: this outside view that “affective neuroscience is slowly settling on a consensus view of suffering” is not happening, and this hurdle to coordination will not resolve itself. Instead, the more affective neuroscience has learned about valence, the more confusing and divergent the picture becomes.
The following is an overview (adapted from my core research) of what affective neuroscience knows about valence.
I’ll front-load some implications for discussion:
- There's lots of philosophical confusion in valence/suffering research. In Kuhnian terms, this would suggest that affective neuroscience is ripe for a paradigm shift. Paradigm shifts often come from outside the field, and usually have unpredictable outcomes (it’s difficult to predict how some future version of affective neuroscience may define suffering).
- Organizations who’ve been using models from affective neuroscience, such as FRI, ACE, and OpenPhil, should be clearer about the caveats involved, and should consider hedging their bets with some ‘basic research’ plays.
- The longer we don’t have a good model for what suffering is, the worse off we’ll be with regard to movement coordination.
Why some things feel better than others: the view from neuroscience
Valence research tends to segregate into two buckets: function and anatomy. The former attempts to provide a description of how valence interacts with thought and behavior, whereas the latter attempts to map valence states to the anatomy of the brain. The following are key highlights from each ‘bucket’:
Valence as a functional component of thought & behavior:
One of the most common views of valence is that it’s the way the brain encodes value:
Emotional feelings (affects) are intrinsic values that inform animals how they are faring in the quest to survive. The various positive affects indicate that animals are returning to “comfort zones” that support survival, and negative affects reflect “discomfort zones” that indicate that animals are in situations that may impair survival. They are ancestral tools for living - evolutionary memories of such importance that they were coded into the genome in rough form (as primary brain processes), which are refined by basic learning mechanisms (secondary processes) as well as by higher-order cognitions/thoughts (tertiary processes). (Panksepp 2010a).
Similarly, valence seems to be a mechanism the brain uses to determine or label salience, or phenomena worth paying attention to (Cooper and Knutson 2008), and to drive reinforcement learning (Bischoff-Grethe et al. 2009).
A common thread in these theories is that valence is entangled with, and perhaps caused by, an appraisal of a situation. Frijda describes this idea as the law of situated meaning: ‘‘Input some event with its particular meaning; out comes an emotion of a particular kind’’ (Frijda 1988). Similarly, Clore et al. phrase this in terms of “The Information Principle”, where “[e]motional feelings provide conscious information from unconscious appraisals of situations.” (Clore, Gasper, and Garvin 2001) Within this framework, positive valence is generally modeled as the result of an outcome being better than expected (Schultz 2015), or a surprising decrease in ‘reward prediction errors’ (RPEs) (Joffily and Coricelli 2013).
Computational affective neuroscience is a relatively new subdiscipline which attempts to formalize this appraisal framework into a unified model of cognitive-emotional-behavioral dynamics. A good example is “Mood as Representation of Momentum” (Eldar et al. 2016), where moods (and valence states) are understood as pre-packaged behavioral and epistemic biases which can be applied to different strategies depending on what kind of ‘reward prediction errors’ are occurring. E.g., if things are going surprisingly well, the brain tries to take advantage of this momentum by shifting into a happier state that is more suited to exploration & exploitation. On the other hand, if things are going surprisingly poorly, the brain shifts into a “hunker-down” mode which conserves resources and options.
However- while these functional descriptions are intuitive, elegant, and appear to explain quite a lot about valence, frustratingly, they fall apart as metaphysically-satisfying answers when we look closely at edge-cases and the anatomy of pain and pleasure.
Valence as a product of neurochemistry & neuroanatomy:
The available neuroanatomical evidence suggests that the above functional themes merely highlight correlations rather than metaphysical truths, and for every functional story about the role of valence, there exist counter-examples. E.g.:
Valence is not the same as value or salience:
(Berridge and Kringelbach 2013) find that “representation [of value] and causation [of pleasure] may actually reflect somewhat separable neuropsychological functions”. Relatedly, (Jensen et al. 2007) note that salience is also handled by different, non-perfectly-overlapping systems in the brain.
Valence should not be thought of in terms of preferences, or reinforcement learning:
Even more interestingly, (Berridge, Robinson, and Aldridge 2009) find that what we call ‘reward’ has three distinct elements in the brain: ‘wanting’, ‘liking’, and ‘learning’, and the neural systems supporting each are each relatively distinct from each other. ‘Wanting’, a.k.a. seeking, seems strongly (though not wholly) dependent upon the mesolimbic dopamine system, whereas ‘liking’, or the actual subjective experience of pleasure, seems to depend upon the opioid, endocannabinoid, and GABA-benzodiazepine neurotransmitter systems, but only within the context of a handful of so-called “hedonic hotspots” (elsewhere, their presence seems to only increase ‘wanting’). With the right interventions disabling each system, it looks like brains can exhibit any permutation of these three: ‘wanting and learning without liking’, ‘wanting and liking without learning’, and so on. Likewise with pain, we can roughly separate the sensory/discriminative component from the affective/motivational component, each of which can be modulated independently (Shriver 2016).
These distinctions between components are empirically significant but not necessarily theoretically crisp: (Berridge and Kringelbach 2013) suggest that the dopamine-mediated, novelty-activated seeking state of mind involves at least some small amount of intrinsic pleasure.
A strong theme in the affective neuroscience literature is that pleasure seems highly linked to certain specialized brain regions / types of circuits:
We note the rewarding properties for all pleasures are likely to be generated by hedonic brain circuits that are distinct from the mediation of other features of the same events (e.g., sensory, cognitive). Thus pleasure is never merely a sensation or a thought, but is instead an additional hedonic gloss generated by the brain via dedicated systems. … Analogous to scattered islands that form a single archipelago, hedonic hotspots are anatomically distributed but interact to form a functional integrated circuit. The circuit obeys control rules that are largely hierarchical and organized into brain levels. Top levels function together as a cooperative heterarchy, so that, for example, multiple unanimous ‘votes’ in favor from simultaneously-participating hotspots in the nucleus accumbens and ventral pallidum are required for opioid stimulation in either forebrain site to enhance ‘liking’ above normal. (Kringelbach and Berridge 2009a)
Some of these ‘hedonic hotspots’ are also implicated in pain, and activity in normally-hedonic regions have been shown to produce an aversive effect under certain psychological conditions, e.g., when threatened or satiated (Berridge and Kringelbach 2013). Furthermore, damage to certain regions of the brain (e.g., the ventral pallidum) in rats changes their reaction toward normally-pleasurable things to active ‘disliking’ (Cromwell and Berridge 1993; Smith et al. 2009). Moreover, certain painkillers such as acetaminophen blunt both pain and pleasure (Durso, Luttrell, and Way 2015). By implication, the circuits or activity patterns that cause pain and pleasure may have similarities not shared with ‘hedonically neutral’ circuits. However, pain does seem to be a slightly more ‘distributed’ phenomenon than pleasure, with fewer regions that consistently contribute.
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.
Finally, the core circuitry implicated in emotions in general, and valence in particular, is highly evolutionarily conserved, and all existing brains seem to generate valence in similar ways: “Cross-species affective neuroscience studies confirm that primary-process emotional feelings are organized within primitive subcortical regions of the brain that are anatomically, neurochemically, and functionally homologous in all mammals that have been studied.” (Panksepp 2010b) Others have indicated the opioid-mediated ‘liking’ reaction may be conserved across an incredibly broad range of brains, from the very complex (humans & other mammals) to the very simple (c. elegans, with 302 neurons), and all known data points in between- e.g., vertebrates, molluscs, crustaceans, and insects (D’iakonova 2001). On the other hand, the role of dopamine may be substantially different, and even behaviorally inverted (associated with negative valence and aversion) in certain invertebrates like insects (Van Swinderen and Andretic 2011) and octopi.
A taxonomy of valence?
How many types of pain and pleasure are there? While neuroscience doesn’t offer a crisp taxonomy, there are some apparent distinctions we can draw from physiological & phenomenological data:
There appear to be at least three general types of physical pain, each associated with a certain profile of ion channel activation: thermal (heat, cold, capsaicin), chemical (lactic acid buildup), and mechanical (punctures, abrasions, etc) (Osteen et al. 2016).
More speculatively, based on a dimensional analysis of psychoactive substances, there appear to be at least three general types of pleasure: ‘fast’ (cocaine, amphetamines), ‘slow’ (morphine), and ‘spiritual’ (LSD, Mescaline, DMT) (Gomez Emilsson 2015).
Mutations in the gene SCN9A can remove the ability to feel any pain mediated by physical nociception (Marković, Janković, and Veselinović 2015; Drenth and Waxman 2007)- however, it appears that this doesn’t impact the ability to feel emotional pain (Heckert 2012).
However, these distinctions between different types of pain & pleasure appear substantially artificial:
Hedonic pleasure, social pleasure, eudaimonic well-being, etc all seem to be manifestations of the same underlying process. (Kringelbach and Berridge 2009b) note: “The available evidence suggests that brain mechanisms involved in fundamental pleasures (food and sexual pleasures) overlap with those for higher-order pleasures (for example, monetary, artistic, musical, altruistic, and transcendent pleasures).” This seems to express a rough neuroscientific consensus (Kashdan, Robert, and King 2008), albeit with some caveats.
Likewise in support of lumping emotional & physical valence together, common painkillers such as acetaminophen help with both physical and social pain (Dewall et al. 2010).
A deeper exploration of the taxonomy of valence is hindered by the fact that the physiologies of pain and pleasure are frustrating inverses of each other.
The core hurdle to understanding pleasure (in contrast to pain) is that there’s no pleasure-specific circuitry analogous to nociceptors, sensors on the periphery of the nervous system which reliably cause pleasure, and whose physiology we can isolate and reverse-engineer.
The core hurdle to understanding pain (in contrast to pleasure) is that there’s only weak and conflicting evidence for pain-specific circuitry analogous to hedonic hotspots, regions deep in the interior of the nervous system which seem to centrally coordinate all pain, and whose physiological mechanics & dynamics we can isolate and reverse-engineer.
I.e., 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.
Philosophical confusion in valence research:
In spite of the progress affective neuroscience continues to make, our current understanding of valence and consciousness is extremely limited, and I offer that the core hurdle for affective neuroscience is philosophical confusion, not mere lack of data. I.e., perhaps our entire approach deserves to be questioned. Several critiques stand out:
Neuroimaging is a poor tool for gathering data:
Much of what we know about valence in the brain has been informed by functional imaging techniques such as fMRI and PET. But neuroscientist Martha Farah notes that these techniques depend upon a very large set of assumptions, and that there’s a widespread worry in neuroscience “that [functional brain] images are more researcher inventions than researcher observations.” (Farah 2014) Farah notes the following flaws:
Neuroimaging is built around indirect and imperfect proxies. Blood flow (which fMRI tracks) and metabolic rates (which PET tracks) are correlated with neural activity, but exactly how and to what extent it’s correlated is unclear, and skeptics abound. Psychologist William Uttal suggests that “fMRI is as distant as the galvanic skin response or pulse rate from cognitive processes.” (Uttal 2011)
The elegant-looking graphics neuroimaging produces are not direct pictures of anything: rather, they involve extensive statistical guesswork and ‘cleaning actions’ by many layers of algorithms. This hidden inferential distance can lead to unwarranted confidence, especially when most models can’t control for differences in brain anatomy.
Neuroimaging tools bias us toward the wrong sorts of explanations. As Uttal puts it, neuroimaging encourages hypotheses “at the wrong (macroscopic) level of analysis rather than the (correct) microscopic level. … we are doing what we can do when we cannot do what we should do.” (Uttal 2011)
Neuroscience’s methods for analyzing data aren’t as good as people think:
There’s a popular belief that if only the above data-gathering problems could be solved, neuroscience would be on firm footing. (Jonas and Kording 2016) attempted to test whether the field is merely data-limited (yet has good methods) in a novel way: by taking a microprocessor (where the ground truth is well-known, and unlimited amounts of arbitrary data can be gathered) and attempting to reverse-engineer it via standard neuroscientific techniques such as lesion studies, whole-processor recordings, pairwise & granger causality, and dimensionality reduction. This should be an easier task than reverse-engineering brain function, yet when they performed this analysis, they found that “the approaches reveal interesting structure in the data but do not meaningfully describe the hierarchy of information processing in the processor. This suggests that current approaches in neuroscience may fall short of producing meaningful models of the brain.” The authors conclude that we don’t understand the brain as well as we think we do, and we’ll need better theories and methods to get there, not just more data.
Subjective experience is hard to study objectively:
Unfortunately, even if we improved our methods for understanding the brain’s computational hierarchy, it will be difficult to translate this into improved knowledge of subjective mental states & properties of experience (such as valence).
In studying consciousness we’ve had to rely on either crude behavioral proxies, or subjective reports of what we’re experiencing. These ‘subjective reports of qualia’ are very low-bandwidth, are of unknown reliability and likely vary in complex, hidden ways across subjects, and as (Tsuchiya et al. 2015) notes, the methodological challenge of gathering them “has biased much of the neural correlates of consciousness (NCC) research away from consciousness and towards neural correlates of perceptual reports”. I.e., if we ask someone to press a button when they have a certain sensation, then measure their brain activity, we’ll often measure the brain activity associated with pressing buttons, rather than the activity associated with the sensation we’re interested in. We can and do attempt to control for this with the addition of ‘no-report’ paradigms, but largely they’re based on the sorts of neuroimaging paradigms critiqued above.
Affective neuroscience has confused goals:
Lisa Barrett (Barrett 2006) goes further and suggests that studying emotions is a particularly hard task for neuroscience, since most emotions are not “natural kinds” i.e.. things whose objective existence makes it possible to discover durable facts about. Instead, Barrett notes, “the natural-kind view of emotion may be the result of an error of arbitrary aggregation. That is, our perceptual processes lead us to aggregate emotional processing into categories that do not necessarily reveal the causal structure of the emotional processing.” As such, many of the terms we use to speak about emotions have only an ad-hoc, fuzzy pseudo-existence, and this significantly undermines the ability of affective neuroscience to standardize on definitions, methods, and goals.
In summary, affective neuroscience suffers from (1) a lack of tools that gather unbiased and functionally-relevant data about the brain, (2) a lack of formal methods which can reconstruct what the brain’s doing and how it’s doing it, (3) epistemological problems interfacing with the subjective nature of consciousness, and (4) an ill-defined goal, as it’s unclear just what it’s attempting to reverse-engineer in the first place.
Fig 1 summarizes some core implications of current neuroscience and philosophical research. In short: valence in the human brain is a complex phenomenon which defies simple description. Affective neuroscience has been hugely useful at illuminating the shape of this complexity, but is running into hugely diminishing returns with its current paradigm, and offers multiple conflicting models of what valence & suffering could be.
Figure 1, core takeaways of affective neuroscience on valence
Cromwell, Howard C., and Kent C. Berridge. 1993. “Where Does Damage Lead to Enhanced Food Aversion: The Ventral Pallidum/substantia Innominata or Lateral Hypothalamus?” Brain Research 624 (1-2): 1–10.
Dewall, C. Nathan, Geoff Macdonald, Gregory D. Webster, Carrie L. Masten, Roy F. Baumeister, Caitlin Powell, David Combs, et al. 2010. “Acetaminophen Reduces Social Pain: Behavioral and Neural Evidence.” Psychological Science 21 (7): 931–37.
Durso, Geoffrey R. O., Andrew Luttrell, and Baldwin M. Way. 2015. “Over-the-Counter Relief From Pains and Pleasures Alike: Acetaminophen Blunts Evaluation Sensitivity to Both Negative and Positive Stimuli.” Psychological Science 26 (6): 750–58.
Jensen, Jimmy, Andrew J. Smith, Matthäus Willeit, Adrian P. Crawley, David J. Mikulis, Irina Vitcu, and Shitij Kapur. 2007. “Separate Brain Regions Code for Salience vs. Valence during Reward Prediction in Humans.” Human Brain Mapping 28 (4): 294–302.
Kashdan, Todd B., Biswas-Diener Robert, and Laura A. King. 2008. “Reconsidering Happiness: The Costs of Distinguishing between Hedonics and Eudaimonia.” The Journal of Positive Psychology 3 (4): 219–33.
Osteen, Jeremiah D., Volker Herzig, John Gilchrist, Joshua J. Emrick, Chuchu Zhang, Xidao Wang, Joel Castro, et al. 2016. “Selective Spider Toxins Reveal a Role for the Nav1.1 Channel in Mechanical Pain.” Nature 534 (7608): 494–99.
Tsuchiya, Naotsugu, Melanie Wilke, Stefan Frässle, and Victor A. F. Lamme. 2015. “No-Report Paradigms: Extracting the True Neural Correlates of Consciousness.” Trends in Cognitive Sciences 19 (12): 757–70.
The above review actually understates the challenge of getting a good model of suffering, because it mostly avoids problems relating to consciousness (of which there are many). Still, my intent here isn't to be discouraging-- or even to throw cold water on the idea that someday, EA could have a good, integrative definition of suffering we could confidently use for animal welfare, AI safety, and social interventions alike. It should be clear from my work that I do think that's possible.
Rather, my point is that EA should be realistic about how bad the current state of knowledge about suffering is, and that this problem isn't going to solve itself.
Mike Johnson, Qualia Research Institute