In 1997, Schultz and colleagues published the elegant and hugely influential theory that argued for the presence of a prediction error in DA neurons, which endowed cues antecedent to rewards with cached value.http://science.sciencemag.org/content/275/5306/1593 …
-
-
Prikaži ovu nit
-
This means that the DA prediction error makes cues paired with rewards valuable in and of themselves, without teaching the subject about the qualitative association between the cue and the reward. [This is the concept inherent in learning driven by the TDRL prediction error.]
Prikaži ovu nit -
However, last year we (and others) showed that the dopamine prediction error COULD function as a teaching signal to drive the development of qualitative cue-cue associations in a manner that transcends this cached-value signal.https://www.nature.com/articles/nn.4538 …
Prikaži ovu nit -
But in this paper we did not test whether the prediction error functions as a cached-value signal. That is, the current state of the literature is that the dopamine prediction error functions as BOTH an associative learning signal AND a cached-value signal.
Prikaži ovu nit -
In this current paper we performed three experiments to test whether the dopamine prediction error acts as a cached-value signal. In each, we found evidence that the error functions to facilitate normal associative learning, WITHOUT endowing cues with scalar value.
Prikaži ovu nit -
In two of our experiments, we again turned to sensory preconditioning (SPC). This entails teaching subjects to associate two neutral auditory cues together. Then, on another day, one of these cues is paired with reward. After this, subjects expect both cues will lead to reward.
Prikaži ovu nit -
We used this procedure in 2/3 exps again, not because we’re lazy
, but because it allows us to stimulate DA while subjects are learning about the cue-cue associations when no rewards are present, so that we can isolate any value given to the cues by DA stimulation.Prikaži ovu nit -
Exp 1: we used the blocking of SPC to reduce cue-cue learning. Then, we showed that DA stim unblocked learning of cue-cue relationships WITHOUT endowing these cues with value. That is, they learnt about the cue, but would not press a lever for it (i.e. conditioned reinforcement).
Prikaži ovu nit -
Exp 2: Usually too many cue-cue pairings reduces SPC. We stimulated DA during normal SPC and found it REDUCED appetitive responding to the DA-paired cue, consistent with over learning. This is in OPPOSITION to a value hypothesis. [and they still would not press for the cues].
Prikaži ovu nit -
Exp 3: we generalised our findings to a context with cues and rewards. Here, we stimulated DA during a configural discrimination: A-no food, AX-food, B-food, BX-no food. Here, DA stim facilitated learning on food AND no food trials, again in OPPOSITION to a value hypothesis.
Prikaži ovu nit -
What’s more is that these subjects STILL would not press a lever more for the DA-paired cues. Importantly, this was after 280 trials of dopamine stimulation!!!
Prikaži ovu nit -
Finally, we showed that our subjects would happily press a lever to get direct stimulation of DA neurons. This showed that we’ve got the right neurons, and we can get behavior that looks like value when delivered outside a context appropriate for a prediction error.
Prikaži ovu nit -
Our conclusion then, is that the dopamine prediction error functions solely as a teaching signal to teach us about the structure of our environment and the associative relationships within it, without making things valuable.
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.