Notes on Neuroscience of human social instincts

Notes/summary:

• Steering subsystem (~amygdala, hypothalamus, brainstem)
  • Innate/genetically specified, so active at birth
  • Houses heuristic mappings from primitive sensory inputs to physiological states
  • Contains “primary rewards,” steers the learning subsystem
  • Ex: reflexive physiological arousal when seeing a spider-like object
• Learning subsystem (~cortex, striatum)
  • Within-lifetime learning, so mostly useless at birth
  • Houses world model
  • Ex: physiological arousal when thinking about a spider
• Symbol grounding problem: how does the brain wire up inputs that don’t exist until later in life (eg cortical concept of a spider) with outputs like physiological arousal?
  • Implementation of human social instincts, among many other things, must solve this problem
  • Ex: how does the brain wire up thinking about a person -> excitement?
• Steering subsystem has heuristics for conspecifics (other humans), termed the “social attention reflex.” When you hear a human vs a bird, you pay special attention
• Learning subsystem contains “short-term predictors,” trained in a supervised way to associate cortical world model concepts with Steering subsystem outputs
  • Context/input = currently active cortical world model concepts
  • Output = estimate of the target Steering subsystem signal
  • Ground truth = output of Steering subsystem (eg physiological arousal)
• Short-term predictors learn to associate cortical world model concepts with active physiological states – every time I go into this particular room I see a spider, so just going into the room triggers the same response before even seeing the spider
• Short-term predictors have a “defer-to-predictor” mode where the output of the predictor can itself trigger the Steering subsystem reaction
  • This seems like a misnomer to me, at least in the example provided in the article? When you’re just thinking about a person, the Steering subsystem isn’t getting visual input of that person, so only the inputs of the learned short-term predictor are active. It’s not really deferring to the predictor, it’s just that only the short-term predictor is active
• A TD learning style algorithm can convert a short-term predictor into a long-term predictor – the predictor can learn to activate the relevant Steering subsystem physiological state at the first cue of a causal chain. Ex:
  • You always only get in your car to go to a restaurant to eat, which feels good
  • Short-term predictor learns that picking up a fork with food in front of you -> feel good
  • Long-term predictor uses a TD-style algo to learn that getting in your car -> feel good
• Recap: predictors learn to associate cortical world model concepts with physiological state over various time delays, with ground truth provided by the Steering subsystem
• How do we ensure that predictors learn useful world model<>physiological state associations?
  • Steering subsystem uses “involuntary attention” to force your cortex to think about what you are seeing/hearing/sensing (instead of eg daydreaming), so that likely relevant cortical world model concepts are active for forming associations
• With everything so far you get the following:
  • Steering subsystem – see another person -> activate social attention reflex
  • When you see another person, Steering subsystem also triggers involuntary attention to force you to think about that person
  • Predictor learns to associate thinking about person¹ -> social attention reflex
  • Defer-to-predictor mode means that the predictor can itself trigger the reflex – purely thinking about that person triggers the social attention reflex
• Another way to see this picture:
  • Predictor is constantly trying to predict what Steering subsystem will do
  • After some initial learning period (how long?), Steering subsystem starts defaulting to defer-to-predictor mode, relying on the learned predictor rather than its own innate heuristics
  • Sometimes, the Steering subsystem injects ground truth when the predictor is wrong for long enough (what are the rules here?), and the predictor updates

Footnotes

1. more precisely, learn to associate “cortical world model concepts that are active when you saw that person” with the social attention reflex ↩