A model of depression (n=1)

My mood at any given time seems to be correlated very tightly to something relating to my output; but naively this would predict that as my capacities increase over time, that I’d get happier, and this mostly doesn’t seem to be the case. A different model I’ve been wondering about is if my mood might be based on the diff from the three-month moving average of my skill/capacity. So when I manage to learn more or get more done than I’m used to managing, I feel awesome; when I seem do be doing or learning less, I feel bad about myself, and hopeless. And of course then that causes me to get even less done, which makes me feel worse… and at some point I hit some kind of rock bottom where some other mechanism is causing me to make a bare minimum amount of progress regardless of how bad I feel. And after a while, that starts to feel normal, and I stop expecting any better from myself… at which point I start having more energy, and manage to do a bit better than I have. Which then causes me to feel actually quite good about myself (since the three-month average has gotten pretty low at this point), which lets me get more done, and so on from there.

This seems to do a pretty good job of predicting the cycling of my low-mood episodes, but doesn’t explain what causes them in the first place. Random fluctuation could explain it, but large situational disruptions to my productivity are relatively unusual, and my mood seems to be less robust than that frequency would predict. So instead of a diff from my three-month average, let’s say that my mood “thermostat” is expecting a linear extrapolation of that three-month dataset, which means that when things have been improving for a while I expect them to keep improving in a linear fashion. And maybe linear growth just can’t be consistently sustained (at least on my current mindware tech, blah blah growth mindset blah blah), and so even minor disruptions can knock me below that day’s mental setpoint for “you’re doing OK at life, you get to feel good today.” And sometimes a counter-fluctuation manages to bring me back up above the setpoint before the positive feedback loop really takes hold, but sometimes not.

The obvious theory for what causes this, if it is basically accurate, is that the guy inside my head who controls my mood thinks his job is to let me know whether I’m doing an OK job, and he’s got some model for how to predict what “an OK job” is on any given day/week. And that his current model is this “linear extrapolation from a 3 month moving average” thingie. However! It’s clear that he’s already sometimes prepared to make exceptions. When my brother-in-law died, even though I wasn’t all that productive in the months afterwards relative to the months leading up to it, I basically didn’t feel bad about myself at all. (I was sad a lot, and angry sometimes, and somewhat less objectively productive as mentioned, so arguably I was something that could be called “depressed,” but there’s something different I want to point at which I clearly wasn’t. In general, I think I often use the word “depressed” to mean something that might be entirely different than what it means clinically; anyone know if this is true, and if so, whether the clinical world has a word for the thing I want?)

Bottom line: it’d be awesome if I was able to have some dialogues with this guy about what sorts of things are sensible “exceptions” to the expectation of linear growth. Or maybe about the linear growth expectation in general, if it’s not actually sustainable. In particular it seems very plausible there’s something like seasonal fluctuations in energy level that I’m not accounting for and am therefore massively accentuating.  Or, maybe this guy is actually fundamentally wrong about what mood is for, and I can convince him of that, instead.

2 thoughts on “A model of depression (n=1)

  1. Related: Mood as Representation of Momentum in a recent Trends in Cognitive Sciences suggests that mood is a computational shortcut that biases our overall expectations upwards or downwards in a situation where there’s a reason to do so (i.e. when recent outcomes have been systematically better/worse than expected), without needing to make a large number of individual belief-updates.

    > Recent research has used experience sampling to examine momentary mood fluctuations during a laboratory-based probabilistic reward task in which monetary rewards varied from trial to trial [26]. The main conclusion of the study was that happiness depends not on how well things are going (in terms of cumulative earnings) but whether they are going better than expected. In particular, self-reported happiness depended on ‘reward prediction errors’ (RPEs; Box 1), that is, the difference between expected outcomes and obtained outcomes. The laboratory results were also replicated in a large-scale smartphone-based experiment with 18420 participants. […]

    > …positive mood induces risk-taking in laboratory experiments [47,48] and in real financial markets [49,50], possibly by biasing upwards the perceived probability of future positive outcomes [51]. In addition, repeated positive RPEs, which should improve mood [26], invigorate reward-seeking behavior [52–55], possibly reflecting an implicit belief in greater reward availability. Furthermore, a positive emotional state reinforces, and a negative emotional state inhibits, one’s current mode of thought, presumably by biasing perception of how well that mode of thought is functioning [56–58]. Finally, many studies suggest that a depressed mood is associated with greater attention or sensitivity to negative information, an effect that may underlie biased perception of outcomes. Notably, both effects can be seen to reflect an implicit belief that things are worse than the objective evidence suggests [59,60].

    > The upshot of this research is that mood induced by a stimulus can affect judgment about other, potentially unrelated, stimuli. Indeed, this property may have given mood its reputation as a rich fountain for irrational behavior. Any attempt to rationalize moods must therefore explain how such biased judgments, which in some cases may reinforce irrelevant actions, nevertheless promote adaptive behavior.

    > According to current theories, agents can maximize reward by keeping track of how much reward is obtained in each experienced state of the environment, and then choosing actions that return them to the states in which such reward has been most abundant [7,8]. For example, an animal using such a mechanism can learn which specific trees bear more fruit and focus its foraging efforts accordingly. This type of ‘reinforcement learning’ algorithm [9] constitutes a powerful way to learn about the environment and converges upon optimal behavioral policies (e. g., [61]). However, there are many real-world situations for which such an algorithm may be poorly equipped. We propose that the information represented by mood is used to mitigate problems that arise in the application of reinforcement learning to such real-world problems.

    > One such learning inefficiency arises when changes in reward in different states are correlated. For instance, increased rainfall or sunshine may cause fruit to become more abundant in all trees simultaneously. In this situation, it makes little sense to update expectations for each tree independently, and a more efficient learning algorithm would instead infer a general increase in reward and update expectations for all related trees accordingly. We suggest this is the function of mood. If fruit becomes more abundant in all trees, a foraging animal will be positively surprised multiple times as it visits adjacent trees and, as a result, its mood will improve. Improved mood will bias the subjective reward for each subsequent fruit upwards, and because these observations are used to update expectations, expectations associated with these trees will be adjusted upwards more rapidly than they would be otherwise. In essence, the effect of positive surprises will be enhanced as more positive surprises are encountered.

    > Through the existence of mood, as an animal learns from experience, its expectations come to reflect not only the reward associated with each particular state (e.g., each tree), but also recent overall changes in the availability of reward in its environment. In this way, learning can account, albeit approximately, for the impact of multiple general environmental factors without having to directly infer the number of factors or the extent of their impact (Box 2). We have described one scenario in which this can be beneficial, but such a generalization mechanism can improve the efficiency of learning in any environment in which different sources of reward are interdependent. Indeed, such interdependencies may be the rule rather than the exception, for both animals and humans, because success in acquiring skills, material resources, social status, and even mating partners can be tightly correlated.

    > Mood can also be useful for learning in another common scenario in which current changes in reward predict later changes in reward. Many processes in the natural world have such momentum. For instance, initial increases in fruit availability may indicate that spring is coming and that further increases are probable. In such a case, a positive mood would represent inference of a positive momentum – which would, in turn, bias perception of subsequent rewards upwards. Because rewards would then be perceived as better than they really are, expectations would be updated upwards quickly and would catch up with rising rewards. Similarly, if reward availability is decreasing in an environment (e.g., winter is coming), then a negative mood leads to rewards being perceived as less good than they actually are (even though increasingly rare rewards still result in positive RPEs) and expectations will catch up with declining rewards, allowing behavior to be quickly adjusted (e.g., hibernate). In accordance with this idea, the relationship between mood and reward perception suggested by the recent literature can be formally derived as statistical inference of average reward and its momentum (Box 2).

Leave a Reply

Your email address will not be published. Required fields are marked *