270 likes | 453 Views
Reward Processing and Social Exchange. Read Montague Baylor College of Medicine Houston, TX www.hnl.bcm.tmc.edu. I love you. Let’s do lunch. A. A. A. B. B. B. Natural visual statistics. Specific algorithms evolved to solve efficiently the array of problems of early vision.
E N D
Reward Processing and Social Exchange Read Montague Baylor College of Medicine Houston, TX www.hnl.bcm.tmc.edu
I love you Let’s do lunch A A A B B B
Natural visual statistics Specific algorithms evolved to solve efficiently the array of problems of early vision
Recharge or die The real world imposes another difficult requirement on all mobile organisms
Natural reward-harvesting statistics? generic woodland creature
What computations should we expect? Reward harvesting is an economic problem involving both valuation and choice Models of reinforcement learning can provide insight
guidance signals Actual experience Counterfactual experience ‘what could have been’ Goal-directed choice Select goal Sustain goal Pursue goal
naive R time burst After learning R time pause After learning (catch trial) time Midbrain dopamine neurons Pause, burst, and ‘no change’ responses represent reward prediction errors – ongoing emission of information ERROR SIGNAL = current reward + g next prediction - current prediction
Can we detect a reward prediction error signal in a human subject?
Passive conditioning tasks Instrumental conditioning tasks Sequential decision tasks Neural correlates of these error signals have been observed in conditioning and decision tasks Social and economic exchange tasks
“I want to solve Fermat’s last theorem” Midbrain dopamine neurons burst and pause responses encode reward prediction errors ERROR SIGNAL = current reward + g next prediction - current prediction Can these systems be re-deployed for abstractly defined rewards (ideas)?
What about the something more abstract like the expression and repayment of trust?
Trust Modeling Must involve risk (uncertainty) Harvesting returns from another agent
I love you Let’s do lunch A A A B B B TRUST
I Agent 1 Agent 2 R Simplifying and quantifying Trust(Berg et al., 1995; Weigelt and Camerer, 1988) Trust is the amount of money a sender sends to a receiver without external enforcement.
A dynamic version of the Trust game (10 rounds) X3 $20 Investor Trustee
I Agent 1 Agent 2 R What is the behavioral signal that most strongly influences changes in trust (money sent) ? Reciprocity = TIT-FOR-TAT
Reciprocity = TIT-FOR-TAT Benevolent signal + - Malevolent signal money sent to partner Neutral signal
Surprise response Questions about brain response Responses that differentiate benevolent from neutral? Responses that differentiate malevolent from neutral? Responses that differentiate benevolent from malevolent?
increases or decreasesin future trust by trustee Trustee will increase trust on next move * Trustee will decrease trust on next move submit reveal time (sec) Reputation develops 0.2 0.2 0 0 * -0.2 -0.2 time (sec) -8 0 10 Trustee Brain: ‘intention to increase trust’ shifts with reputation building Signal is now anticipating the outcome
naive R time burst After learning R time pause After learning (catch trial) time Temporal shift resembles value transfer in reward learning experiments
Why should intentions to increase repayment burn more energy? 1 0 .9 -.1 .8 -.2 .7 -.3 TRUSTEE increases repayment .6 TRUSTEE decreases repayment -.4 .5 -.5 .4 -.6 -.7 .3 -.8 .2 -.9 .1 r = .00 r = .27 -1 0 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Change in next investment Change in next investment Trustee decreases No information Trustee decreases positive correlation
Trust? • Social trust is about modeling others - always some underlying currency. • Re-deploy reward-harvesting machinery that we share with all vertebrates (abstractions gain reward status) • Use to probe pathologies (addicted state, autism spectrum disorder, borderline…)
Collaborators Baylor College of Medicine Pearl Chiu Amin Kayali Brooks King-Casas Terry Lohrenz Sam McClure (Princeton) Read Montague Damon Tomlin Caltech Cedric Anen Colin Camerer Steve Quartz UCL Peter Dayan Nathaniel Daw Salk Institute Terry Sejnowski Princeton University Jon Cohen Emory University Greg Berns University of Alabama Laura Klinger Mark Klinger Families of autistic subjects Funding Sources NIDA NIMH Dana Foundation Kane Family Foundation Angel Williamson Imaging Center Institute for Advanced Study, Princeton NJ www.hnl.bcm.tmc.edu/trust
Investment phase cue to invest investperiod delayperiod investment revealed to both brains . . . 4 s 8 s 10 s cue to guess guessperiod synchronized on submission time of both partners .4 .3 fraction of highly accurate guesses by trustee .2 .1 3 4 5 6 7 8 9 10 rounds How do we know a reputation is forming? Investor events Trustee events Trustee
i Split the profit i Split ‘loaf’ i What is Fair? Trustee Investor 20 -i i Split the total 10 + i each collective ownership? common goods?