Deep down, mood is not about feeling.
Cognitive behavioral theory, behavioral ecology and computational psychiatry agree: The crux of mood is expectation, not feeling.
If The Onion would ever care to satirize computational psychiatry, it would run a headline like: “Groundbreaking Research Confirms Depressed People Expect Things Will Go Badly". But they won't - academic papers like What is Mood? A Computational Perspective 1 aren’t exactly clickbait - the article I’ll be explicating today! It is dense and abstract, and one might think that it can’t tell us anything about the lived experience of mood. Computation and mood indeed! Isn’t it ridiculous to offer a computational perspective on that innermost, ineffable experience?
Well, you be the judge. I found that a computational perspective on mood, by going deep where mood blends with predictive processing and information theory, offers conceptual reshufflings, clinical pearls and insights into the phenomenological mysteries of mood experience.
So you can see that I was just teasing computational psychiatry. While it’s true that the computational perspective partly restates what we already know - that depressed people expect things to go badly - it offers unforeseen vantage points from which to contemplate mood. For instance, positing that what is utterly essential to mood is not “feeling” but world-interpreting - in the case of depressed mood, for instance, that “expecting things to go badly” amounts to elevated certainty that the results of action will be uncertain. I did not garble that last statement! We will be wrestling indeed with mind-twisters like “the certainty of uncertainty”.
While feeling may be a universal feature of mood and mood disorders, evolutionary and computational perspectives suggest that altered predictive processing better carves nature at its joints. If this is true, diagnostic manuals like DSM-5-TR have misplaced their emphasis by elevating subjective feelings (such as "feeling blue" or "elevated") as essential diagnostic criteria for mood episodes. These subjective experiences are "Criterion A" symptoms required for a persistent mental state to qualify as a mood episode, when in fact the core disturbance lies in prediction and certainty about the consequences of action.
Allow me to emphasize a point that we will return to later:
DSM-5-TR misplaces its emphasis by treating subjective feelings (such as "feeling blue" or "elevated") as essential diagnostic criteria for mood episodes.
Rather than tackle the entire article - let alone computational psychiatry as a whole - I'll avoid biting off more than I can chew and focus on how it defines the fundamental nature of emotion and mood.
Three streams converge
Three streams of thought have converged on the kernel of mood. The cognitive model of depression first highlighted how "thoughts" (really, world-interpreting tendencies) are central to clinical depression. Theoretical behavioral ecology then proposed that mood was about prospectively fine-tuning the thresholds for emotions (see my post Psychiatrists who don't know what mood is). This seemed like the absolute rock-bottom, evolutionary foundation of what adaptationist and phylogenetic biology could theorize about mood. Well, along comes computational psychiatry, exemplified in our article, daring to dig deeper. Result: the three streams converge on how mood's primary function is to bias cognition and judgment, tuning how we predict and interpret the world.
Caveat: I’m no expert in computation (though I did manage to get a B+ in college calculus 47 years ago…) I’m just a scout from the realm of clinical psychiatry venturing into the realm of computational psychiatry and evolutionary theory, realms that are rife with Chasms of Incomprehensibility. You’ll discover interesting vistas if you follow me, but keep an eye on your own GPS as we go.
So let’s take a look at What is mood? A computational perspective. In addition to those interesting vistas, we will glean some clinical pearls, clarify psychiatric diagnosis, and ultimately, illuminate our understanding of what it means to be a sentient being.
Normal mood variation
As we venture into this disorienting landscape, let's get our bearings contemplating how mood relates to information. After all, moods are set off by the information we receive, and are sustained by how we continue to process that information. This gives us our first clue about the “computational' perspective” - the systematic processing of information.
Let’s proceed to unpack how the authors Clark, Watson and Friston understand emotions. Now, don’t get your hopes up - their theory does not explain the richness and variety of emotions. The authors are in line with Lisa Feldman Barrett’s theory of constructed emotion, in which the complexity of any instance of emotion is assembled on the fly from a panoply of memories, conceptual schemas and interpretation of bodily signals (interoception). When Clark, Watson and Friston say “emotion” they are merely referring to its valence - the overall “good vs. bad” feeling they contain. What are the computations that determine which emotions will be experienced as good, and which as bad?
Keeping in mind the modest scope of their claim, here is how they identify emotional states:
"Emotional states reflect changes in the uncertainty about the somatic consequences of action".
It goes without saying changes are what trigger emotions, but notice that the changes in question here are not “out in the world” per se, like hitting the jackpot or getting that dreaded "We need to talk" text from your significant other. The changes the authors refer to occur downstream from those of the outer world - changes in something unconscious and computational - namely, “uncertainty”.
What kind of events (changes) in the world induce downstream changes in the “uncertainty about the consequences of action”? Those are changes (events) that undermine the precision of prior beliefs about the efficacy of action.
Two clarifications: Firstly, beliefs (in this context) are not necessarily conscious propositions; they could be unconscious expectations. Secondly, the technical meaning of precision refers to an unconscious, probabilistic estimation concerning the certainty of expectations. Precision, in this technical sense, is the inverse of uncertainty.
Let’s flesh this out with an example of an emotion that carries a negative valence. Imagine you’re going about your day when something happens that contradicts your expectations. At first, you may experience surprise - which may well be the subjective correlate of prediction error. But this may indicate that your actions are not resulting in their expected consequences - that you are losing your grip on the world.
And your brain cares deeply about keeping track of your grip! It must calibrate the extent of your agency - of your power and efficacy. So it lowers the estimated precision of action-outcome predictions. As well it should, if it’s going to accurately model a certain entity-in-its-context - the entity known as you.
The crux of the matter that the brain (the generative model) cares about is, how likely is it that acting will make the difference you predict it will ? If there are events indicating lowering of this probability, there will follow a drop in precision. And this is subjectively experienced as negatively valenced -as the negative valence component of whatever emotion your brain constructs. Helplessness and uncertainty about the consequences of action, again, a feeling that you are losing your grip on the world.
“… negative emotions contextualize events that induce expectations of unpredictability, while positive emotions refer to events that resolve uncertainty and confer a feeling of control”.
So much for the gross valence of emotions. How does this join up with understanding mood?
From emotion to mood
Before we get to the technical distinction between emotion and mood proposed by the computational perspective, take a moment to notice that we commonly use these terms loosely. Even professionals whose job entails accurately naming mental states! This should be a basic professional competence of mental health professionals, but it received little attention during my residency training. I recall no more sophisticated guide for distinguishing emotions from moods than by duration, relying on the metaphor “Emotions are like the weather, mood is like the climate”.
The metaphor is apt up to a point… Like the climate, moods are long-lasting and shape what emotions (what weather) is statistically likely to occur. So far so good. But the weather/climate metaphor misses the most important feature of the distinction: Moods aren’t simply long-lasting emotions, but include predictive and interpretive processes - they are prospective, not reactive.
Does the computational perspective help us better pinpoint the difference between emotions and moods? Let’s see how Clark et al formally characterize mood:
"Mood corresponds to hyperpriors about emotional states, or confidence about the consequences of action. In other words, mood states reflect the prior expectation about precision that nuances (emotional) fluctuations in confidence or uncertainty”
Hierarchies and hyperpriors
We just saw how the definition of mood introduced the notion of hyperpriors. Recall that the hierarchical nature of predictive processing requires higher-level predictions that shape how lower levels process information. Just as each level makes predictions about the level below, kind of like executive oversight, predicting the likely patterns in lower-level processing.
This hierarchical system applies to all cognition, and the processing of emotion-related phenomena is no exception. If emotions reflect the precision of prior beliefs about the consequences of action, our brains would be expected to also encode hyperpriors that constrain how short-term emotional states fluctuate, setting meta-level expectations about how confident or uncertain we should be about predicted action outcomes. We would expect to have hyperpriors about emotions.
Besides efficient “oversight”, there is another rationale for the necessity of hierarchical hyperpriors: accounting for time, for duration, for the temporally extended and prospective nature of mood. Hyperpriors are computationally needed to provide temporal depth . Only a processing hierarchy has the power to capture long-term patterns and make predictions over different time scales. Without hyperpriors and hierarchical processing, we would only be able to react to the now with knee-jerk emotions. We’d be flying blind.
Two-dimensional mood
Just like the adaptationist, behavioral ecology article I discussed in the post I mentioned earlier (Psychiatrists who don't know what mood is), the computational perspective maps mood onto a two-dimensional space. But while the earlier theory plotted mood along axes of sensitivity to possible reward and sensitivity to possible punishment, here the dimensions are the statistical mean of prior expectations about the consequences of action and the precision those priors.
Let’s unpack this a bit, using the graph, and take the case of severely depressed mood. The mean value of expectations about how predictably effective our actions will be is low, mapping onto negative values on the y-axis - the line goes down! On the x-axis, we see that such expectations are endowed with high precision (certainty), mapping onto high values - the line goes to the right.
Consider now elevated mood. The blue line representing it is a mirror image of the line of depression - it’s flipped across the x-axis. Elevated mood shows high expectations of agency (high mean on the y-axis) held with high precision (far right on x-axis).
In mania, this precision about expected control is extreme, as anyone who’s dealt with a manic friend or patient or has personally experienced mania well knows. The uncanny certainty of mania brooks no contradiction. Even psychiatric guidelines advise clinicians to not gainsay it: “…the psychiatrist cannot ‘talk’ a patient out of acute mania.2
This figure is from our article (other than the emojis, which are mine…) showing mood characterized by two parameters: expected predictability (μ, traditionally used in statistics to represent the mean) and certainty about those expectations (τ, representing precision).
Depression (red lines) combines low expectations about predictability (μD) with high certainty about this unpredictability (τD) - in other words, being sure of being helpless. Mania (blue lines) shows high expectations of predictability (μM) with high certainty (τM) - being sure one can control outcomes. Anxiety (green lines) combines low expectations of predictability (μA) and less certainty (lower τA) - leading to frequent attempts to resolve uncertainty.
Clearing up objections
Let’s consider possible objections to this whole framework, a framework that is so alien to our everyday psychology. I’ll present a sample of my own objections as a naive student, and how I attempted to muddle through them playing professor.
Objection #1: "Professor, I’m confused…When I'm depressed, I expect things to go wrong. When I'm happy, I expect things to go well . What’s this about degrees of confidence in control?
We have seen by now how the computational perspective of mood, illustrated in the graph, posits that mood forecasts not “good vs. bad worlds” - not adversity vs. bonanzas - but expectations of how much one will be able to predict and control outcomes.
Can this be right? Well, consider how we qualify the significance of events as good or bad without reflecting on how their impact “plays out” affecting the control/predictability of our world. Imagine a random selection of vicissitudes and life events that you think would result in positive vs. negative emotions. Then play out the life-consequences of those events and ask yourself whether they do not indeed “… induce expectations of unpredictability” (loss of control) or, conversely, “… resolve uncertainty and confer a feeling of control”.
Besides introspective exercises such as this one, there is a line of animal research that supports the link between valence/mood and predictive control, namely, the “stressor controllability paradigm” In contrast to older “learned helplessness paradigm” where laboratory animals were subjected to uncontrollable stressors, the “stressor controllability paradigm” compares animals that receive identical stressors, only one can exercise control over the stressor (its termination), while the other one has no control. Despite experiencing identical “bad events” (stressors), their outcomes differ. In studies using rats, the rat that lacks control develops signs of depression, whereas the rat that can do something expecting it to work is “resilient” to adversity - it does not “get depressed”.
We could ask the rat-with-the-wheel if it expected “things to go keep on going wrong” and it would answer, “Of course I expected things to go wrong - those shocks just keep on coming!” But it had confidence (high precision) that it could control them. The rat with no control experienced, as it were, losing its grip on the world, and this was what made the difference. 3
Objection #2: “Professor, I was just hanging around, doing nothing, when my boyfriend texted that he was dumping me. I cried. Why does this theory say that emotions are always about action, when I was just sitting around when I heard the news?
Indeed, this theory specifies that emotions follow action. And yes, we surely experience emotions after “things happen” that weren’t the “consequences of action - such as out-of-the-blue events. But consider that we are always doing something or other. Whatever our behavioral state was before some out-of-the-blue event may simply compute as its antecedent behavior (action). The emotional system, as it were, chides us: “Be better prepared next time.”
Objection # 3: Professor, I have a Free Energy Principle objection… How can events that are unexpectedly “good” result in positively valenced emotions, if they are surprising?
About those welcome surprises… this seems to contradict the fundamental free-energy principle that organisms strive to minimize surprise and prediction errors. Shouldn’t surprises elicit negative valence, given that prediction errors connote unpredictability and decreased confidence in our model of the world?
Positively valenced emotions, indeed, are associated with increases in the precision of predictions about the future - that is, about the consequences of action. When does it make sense for the precision of predictions to increase? When events confirm that we have a good grip. Or better yet, when events surprise us that our grip is improving, and events "resolve uncertainty and confer a feeling of control”. When this happens, our emotional system seems to conspiratorially whisper in our ear: “You got this!”
The paradox resolves when we consider that like all sentient beings, we have “built-in” fundamental priors specifying our preferred states. We assume that such states are likely to occur - they are “set” to be less surprising.
Think of it as implicit optimism bias concerning our capacity to maintain homeostasis. These “built-in” default priors about preferred states function as predictions that actions will make a difference, that we are capable agents who have power to influence our environment.
To be alive is to be optimistic!
To be alive is to be optimistic.
In a future post I will tackle this fascinating consequence of active inference - what Karl Friston and colleagues call “self-evidencing” - gathering evidence for our own existence as viable organisms, and the optimistic bias any agent needs. Descarte’s “I think therefore I am” may be superannuated! Deeper than the solipsistic chauvinism of “thinking” lies the dictum of life: My actions work as I expect, therefore I am.
“My actions work as I expect, therefore I am”
Objection #4: “Professor, this is all so utterly reductionist. What about our innermost feelings - the subjective phenomenology of mood?”
My dear literary friends, and those with a fondness for Continental philosophy, don’t worry - I do not believe that computational and evolutionary accounts exhaust the significance of mood! I love a good reductionist explanation, but I recognize its limits. It’s just that I’m searching for a primitive nugget, a first-principle, a subatomic particle of mood.
By all means, we should explore the grand expanse of mood’s phenomenology and existential significance - a project that beckons me for some other day. The phenomenology of mood manages to be at once of innermost particularity and of panoramic world-interpreting. What could be of greater interest, especially for psychotherapists who contend with mood disorders? We are fated to have mood states toss us about, much as Heidegger says that we are “thrown” into Dasein. We find ourselves "always already” interpreting the world, including our selves, through the tinted glasses of mood.
But wait! Why am I lapsing into Heidegger-speak when I can’t stand this philosopher’s humorlessness and scholastic obscurity? I must be harboring some deep intellectual contradiction! So, I resolve to face it head on in a post, where I’ll rant about Heidegger and the jargon of postmodernity. Stay tuned…
What then shall we do?
Though we have examined only a narrow slice of computational psychiatry, even this limited view yields some interesting implications. It suggests restructuring the classification of mood disorders, it hints at useful refinements for clinical practice, and provides insights into how our emotional lives reflect our sense of agency and efficacy. Let's consider each in turn.
Updating psychiatric nosology: What is the essence of mood episodes?
It’s old hat by now that the entire enterprise of classifying psychiatric conditions is in bad shape. The prevailing scheme, DSM-5-TR, is bashed over and over for not “carving nature at the joints”, for relying on pick-from-the-menu diagnostic algorithms, for having concocted diagnoses merely to justify drug company sales, for pathologizing variant ways of being, and so on.
Yet in its early days DSM was praised precisely for being a-theoretical - for having divorced itself from speculative psychoanalysis and primitive biological psychiatry. That was a good move in its day! But I’d say it’s time to tentatively incorporate a little theory back into psychiatric classification.
Consider the implications of the computational perspective for the diagnosis of “Generalized Anxiety Disorder”. No doubt about it - it should be re-classified as a mood disorder. Consider how its core feature, "apprehensive expectation" , is fundamentally about prediction. And as with depression and mania, it involves a persistent alteration in interpreting and responding to the world. Furthermore, as we saw in the “mood graph”, anxious mood can be decomposed into the same fundamental dimensions of expected predictability and precision that characterize depressed and manic moods.
But wait…such a reclassification would bring up another problem with DSM’s theoretical agnosticism: it no longer features a superordinate category of “Mood Disorders”! ( The last version to use this umbrella category was DSM-IV-TR, published in 2000). There were reasons for eliminating it, which I won’t get into here. Yet we’ve seen how theoretical behavioral ecology and the computational perspective suggest that mood is a natural kind, which is a strong rationale for future versions of DSM to resurrect the category.
Another implication of the computational perspective is how it shines a spotlight on a diagnostic elephant-in-the-room: The essential feature of mood disorder episodes, namely their cognitive distortions and distorted predictions, is underweighted in their diagnostic criteria. Perhaps because they are not symptoms per se, and the committees who came up with criteria were aspiring to behaviorist purity: Just the symptoms, ma’am. And as the skewed information processing characteristic of clinical depression, mania and anxiety are not symptoms - symptoms are what patients report, but those in the throes of mood disorder episodes are caught within their predictive frameworks, blind to the skewed information processing.
Because patients will never complain of patterns of world-interpreting - those patterns can only be inferred from patients’ stream of thought and behavior. They may also be sussed out with psychometric instruments, such as depression inventories. For example, in the case of depression, the Beck Depression Inventory, one of the classic assessment tools, does a better job than the Hamilton Depression Rating Scale (HAM-D) or the Montgomery-Åsberg Depression Rating Scale (MADRS). But as it stands, even when dysfunctional cognition is evident, it hardly counts towards diagnosis.
Lastly, the predictive processing framework points to a paradox in mood research. While clinicians define mood and mood disorders by their persistence (in contrast to fleeting emotions), some researchers routinely use Experience Sampling Methods that are supposed to track momentary mood. But moods are defined as long-lasting - like climate! What sense does it make to assess mood several times a day - whenever some smart phone app beeps and prompts study subjects to rate their “mood”?
Perhaps the referent of our common usage of the word “mood” bifurcates, and we should come up with distinctive terms for the climate-like, long duration affective phenomenon of mood and another term for… something that iridescently varies moment to moment. Whatever we choose to call “momentary mood”, we could define it as sharing with “climate-mood” the fundamental property of being about prediction, but short-term predictions that change as our sense agency changes throughout the day.
Implications for the treatment of mood disorders
It would be a nice surprise if the computational perspective could offer insights into the biological treatment of mood disorders. But not so far… How could it, when there is no consensus about classification, etiology (fundamental causes), or pathophysiology (dysfunctional mechanistic pathways)? Now, there have been "proof of principle" efforts to connect "the encoding of uncertainty” to "neuromodulatory antidepressants" in some articles (for example, Why depressed mood is adaptive: a numerical proof of principle for an evolutionary systems theory of depression. )Promising, but we are still far from practical biological interventions. 4
All that being said, shedding light on the core features of mood may help us to untangle the mess of conceptual knots in the current literature about mood disorders.
Psychotherapeutically, the computational perspective’s redefinition of emotions and moods (remembering we’re examining just one slice of the theory) may have implication for how clinicians talk with patients.
The consequence of high precision in the case of severely depressed mood is that when prediction errors arise- let’s say one experiences being more effective than expected, generating prediction error - those prediction errors will be squelched. In one’s subjective mentation this might be experienced as “explaining away” the significance of the prediction error.
We know that entreating a seriously depressed friend or patient with reasons to cheer up usually meets with resistance. The computational perspective on such resistance, we have seen, is reflected in how mood may be “decomposed” into the mean of expected predictions and their precision (the inverse of the variance). The resistance is that elevated value on the x-axis.
This maps nicely onto cognitive behavioral therapy (CBT) for depression, particularly the cognitive distortion known as "disqualifying the positive." While typically listed alongside other “negative thinking”, this distortion stands out as the most transparent example of how depression rationalizes its world-view. I’d bet that focusing the therapeutic effort on this particular distortion - understood anew as the active squelching of prediction errors - might be especially effective.
Here are mnemonics for therapists if they wish to make use of the computational perspectives’ three dysfunctional quadrants of mood:
Depressed patients are certain of uncertainty: “I have a poor grip on the world, for sure.”).
(“Depression occurs when an uncertain, unpredictable outcome is predicted with high precision.”)
Manic patients are certain of certainty: “I have a great grip on the world, for sure!”
(“Mania… is characterised by high precision, but with the expectation of a predictable and controllable outcome.”)
Anxious patients are uncertain of uncertainty: “I have a poor grip on the world - or do I?”
(“Anxiety is an expected unpredictability but with low precision. As such, the individual engages in behaviour designed to resolve this uncertainty… which never does.”)
Takeaways and insights
Let's close by exploring the deepest implication - or, if I may mix metaphors, by gazing at the grandest vistas - offered by the computational perspective.
We have seen three roads converge; the CBT road, which has long emphasized the predictive, prospective, world-interpreting essence of mood; the computational road postulating that emotion and mood are fundamentally about confidence in the predictability of our agency, and the behavioral ecology road that focuses on animal behavior and adaptation.
The latter two roads reveal how mood mirrors the world momentum, and is in effect our internal model of it. How the world tends to keep flowing the way it’s been flowing. How both the states of the outer world and the inner world autocorrelate between time t and time t+1. And the computational and ecological roads suggest something else: mood is a natural kind, just as much as forelimb is a natural kind that we can discern across tetrapods - be they humans, horses, bats, whales, birds or pterosaurs.
This computational redescription of emotion and mood offers psychiatry a chance to reconceptualize its diagnostic framework. Perhaps a diagnostic manual that indulges again in theory (but please, 21st century theory) would lead to more effective treatments for those who suffer from mood disorders.
And what about the ordinary fluctuations of mood - their normal ups and downs? By understanding mood in this new way, we gain insights that may boost mastery about our own mental states, when such is called for. If emotions and moods are “just another construct…that the brain employs to deploy precision”5 and if knowledge is power and power increases precision, we may gain a superordinate power – a capacity to have meta-emotions about emotions and meta-moods about moods. This is a possibility only dimly perceived, a faint glimmer… a half-baked hope.
Afterword: Nietzsches’ flash of insight
Nietzsche offers an observation in The Antichrist that illuminates the realms of mood like a lightning flash. In answer to his own question 'What is happiness?' he writes: 'The feeling that power increases - that a resistance is overcome.'
“What is happiness? The feeling that power increases- that a resistance is overcome.”
Please hygienically extract this crystalline Nietschean aperçu from his metaphysical baggage about Will to Power! Take it strictly as a proposition about our animal psychology, and it captures the very essence the computational perspective on mood.
Clark JE, Watson S, Friston KJ. What is mood? A computational perspective. Psychological Medicine. 2018;48(14):2277-2284. doi:10.1017/S0033291718000430
David Kahn, The Psychotherapy of Mania, Psychiatric Clinics of North America,Volume 13, Issue 2, 1990, Pages 229-240,
Maier, S. F., Amat, J., Baratta, M. V., Paul, E., & Watkins, L. R. (2006). Behavioral control, the medial prefrontal cortex, and resilience. Dialogues in clinical neuroscience, 8(4), 397-406.
Constant, A., Hesp, C., Davey, C. G., Friston, K. J., & Badcock, P. B. (2021). Why depressed mood is adaptive: a numerical proof of principle for an evolutionary systems theory of depression. Computational Psychiatry, 5(1), 60.
Ibid.