this article continues the discussion from does ai work like ur brain? where i introduce the bayesian brain model. recommend reading that first if u havent.
in the first article, the central idea was that the brain works like a probabilistic prediction engine. it doesnt passively record the world, it makes constant bets about what will happen and adjusts the model when it gets it wrong.
an interesting consequence of this is, if the brain is a prediction machine, what happens when that machine has a bug?
depression, autism and adhd can be understood exactly like that. not as "character defects" nor as inexplicable mysteries, but as variations in the prediction engine, each with its own internal logic. (⊙_⊙)
the frameworks here are developing models, not consolidated clinical consensus. bayesian theory of depression is influential but its one model among many. same for autism. adhd is even more speculative. im treating each one as what it is, a useful lens for thinking about the problem, not the definitive explanation. diagnosis and follow-up are still with professionals.
depression: priors stuck in negative
in the bayesian framework, priors are the prior beliefs the brain uses to interpret everything that comes in. normally, when new evidence shows up, it updates those beliefs. thats learning happening.
in depression, the priors get too heavy and too negative to be moved.
the depressed brain operates with something like "i will fail", "no one likes me", "nothing will work out". contrary evidence arrives, but doesnt have enough force to move those beliefs. the internal model discards positive data as noise.
u pass a hard exam and the brain says "it was luck". a friend compliments u and u think "theyre being polite". anything good that happens gets filtered out bc it doesnt match the internal model. (¬_¬)
Paul Dayan and other researchers linked this to learned helplessness. if the brain learns that its actions dont influence any outcome, it stops trying to predict a better future. "reward expectation" gets flattened. thats why anhedonia (loss of pleasure) makes sense in this model. the brain stopped predicting that anything will be good.
fristons perverse strategy
remember Friston said the brains goal is to minimize surprise? in depression this turns into something pretty dark.
if i predict everything will go wrong, nothing surprises me. lowering expectations to the max is a form of minimizing prediction error. the brain is technically "optimizing", but it reached a horrible and stable local equilibrium. (x_x)
any attempt to leave this state generates surprise, which means high prediction error, which means the brain resists. thats why "just cheer up" doesnt work. the system is stuck in a local minimum and any movement outwards feels like a threat to the model.
rumination = endless loop
that thing of chewing the same bad thought for hours? researchers like Barrett and Paulus see this as the brain trying to solve a prediction error that has no accessible solution in the current model.
it runs the model again, again and again trying to find coherence that doesnt come. its a processing loop stuck with no stop condition. (¬_¬)
why psychedelics seem to work
this is the part that fascinated me the most.
Robin Carhart-Harris proposed the REBUS theory (Relaxed Beliefs Under Psychedelics). the idea is that psilocybin and LSD temporarily "loosen" super rigid priors, leaving the brain more open to updating the stuck beliefs. like a forced reset of the model. (✪ω✪)
clinical trials with psilocybin for treatment-resistant depression are showing promising results exactly bc the mechanism makes sense inside this framework. its not magic, its decompression of priors that were too heavy to move on their own.
Cognitive Behavioral Therapy can be seen the same way. u literally train the brain to collect new evidence and force belief updates. slower than psilocybin, but the mechanism is analogous. (｡•̀ᴗ-)✧
depression also has genetic, neurochemical (serotonin, dopamine, glutamate), inflammatory, social and traumatic components. the bayesian framework doesnt replace any of that, it offers an elegant mathematical way to understand how those factors manifest in cognition and behavior. if u or someone close is going through this, therapy and medical follow-up are the way.
autism: weak priors, world at max volume
in the typical brain, priors compress reality. u see a chair even if the lighting changes, even if its partially covered by a bag, bc the prior "chair" fills the gaps automatically. the brain trusts the internal model a lot and discards small sensory variations as noise.
the most influential theory about autism, proposed by Pellicano and Burr (2012), is the hypo-priors one. weaker priors relative to sensory evidence. raw sensory input has more weight. the internal model has less weight. (⊙_⊙)
this explains a lot at once:
- sensory hypersensitivity. clothing tag, fluorescent light, food texture. the brain doesnt filter those stimuli as "expected noise". they arrive at max volume bc the prior filter is weaker. >_<
- difficulty with routine changes. if the internal model is less reliable, any change in the environment generates high prediction error, which translates to overload.
- attention to detail. u dont compress the scene as much, so u notice things others filter out.
- social difficulty. inferring intention, sarcasm and subtext depends on strong priors about human behavior. with weaker priors, this takes way more conscious processing.
- stimming and preference for predictable routines. self-regulation. u seek environments where prediction error is low and controllable.
Van de Cruys proposed the HIPPEA hypothesis (High, Inflexible Precision of Prediction Errors). its not that priors are weak, its that the brain treats every prediction error as critical, without being able to distinguish well between noise and relevant signal. the behavioral result ends up being similar, but the mechanism is different.
one point i think is important. this framework reframes autism not as a "deficit", but as a different computational style with tradeoffs. worse at quickly inferring implicit social context. better at detecting patterns and details that go over most peoples heads. (｡•̀ᴗ-)✧
adhd: dopamine and the future that doesnt exist
dopamine in the brain isnt "pleasure" like most people think. it encodes reward prediction error, the famous work by Wolfram Schultz with monkeys in the 90s. when something is better than expected, dopamine spike. worse than expected, drop. when it matches prediction exactly, nothing happens.
in adhd, the hypothesis is that this signal is noisier, or has a steeper temporal discount. distant rewards in time lose value way too fast.
this explains several classic behaviors:
- difficulty with long tasks. the brain literally cant "feel" the reward of finishing the report. its too far away, the prediction signal is too weak to motivate now. (¬_¬)
- novelty and stimulation seeking. u need constant prediction error to keep the system engaged. predictable tasks dont generate enough dopamine.
- hyperfocus. when something generates continuous flow of interesting prediction errors, videogame, new hobby, fascinating project, the system locks in there bc its finally getting the signal it needs.
- impulsivity. if the future is too discounted, immediate small reward beats big future reward. its the rational choice of the system given its time model.
- time blindness. subjective time perception is tied to those predictions. without them calibrated, "now" and "not now" is the only division that really works.
Ritalin and Vyvanse increase available dopamine and norepinephrine. in the bayesian framework this means increasing the precision of the prediction error signal. distant rewards regain enough weight to guide behavior. the brain can better distinguish what matters from noise. (✿◡‿◡)
theres also the hypothesis that the Default Mode Network doesnt turn off properly when u need to focus. the result is the classic. rumination, daydreaming and intrusive thoughts invading the current task at the worst possible moment.
AuDHD: stuck between two opposites
autism and adhd coexist with pretty high frequency. estimates vary but stay around 30-80% overlap depending on the study. at first glance it seems contradictory:
- autism: weak priors, seeks predictability to reduce sensory overload
- adhd: muffled dopamine signal, needs novelty to function
but it makes sense at the same time. the autistic brain wants predictable environments bc weak priors generate overload in chaotic environments. the adhd brain needs novelty bc the dopamine signal is muffled and routine doesnt generate enough stimulus.
result: u get stuck between needing structure to not freak out and needing novelty to not die of boredom inside it. (-_-) hard life.
these models are mathematically elegant but still in validation. heterogeneity within each of these conditions is huge. two autistic people can have completely different profiles. adhd and autism arent "bugs" in the pejorative sense, theyre neurological variations with advantages and disadvantages that depend a lot on context and environment. diagnosis and support are with professionals. these frameworks are conceptual, not self-assessment manuals. (｡•ᵕ•｡)
what i find most valuable in this whole framework isnt the technical explanation itself. its that it removes moral judgment from these conditions. depression isnt weakness. autism isnt defect. adhd isnt lack of willpower. theyre different configurations of a prediction engine running on biological hardware with its own constraints and tradeoffs.
understanding the mechanism doesnt solve everything. but it changes the conversation. (｡•̀ᴗ-)✧
← read part 1? does ai work like ur brain?