Dopamine, Uncertainty and TD Learning CoSyNe’04 Yael Niv Michael Duff Peter Dayan.

Dopamine, Uncertainty

and TD Learning

CoSyNe’04

Yael Niv Michael DuffPeter Dayan

What does Dopamine encode?• Important neuromodulator

- Neurological/psychiatric disorders - Drug addiction/self stimulation

• Fundamental role in RL- Classical/Pavlovian conditioning- Instrumental/operant conditioning

• DA neurons respond to:− Unexpected (appetitive) rewards− Stimuli predicting (appetitive) rewards− Withdrawal of expected rewards− Novel/Salient stimuli

What does Dopamine encode?

DA represents some aspect of reward, but not rewards as such.

The TD Hypothesis of Dopamine

)()(ˆ

)(ˆ)1(ˆ)1()(

ttV

tVtVtrt

)1()1()(

)()(

tVtrtV

rtVt

DA encodes the reward prediction error

<-DA

Stimulus Reward Stimulus RewardStimulus Reward

DA

δ(t)

Precise theory for the generation of DA firing patternsCompelling account for the role of DA in classical conditioning

But: Fiorillo, Tobler & Schultz 2003• Introduce inherent uncertainty into the

classical conditioning paradigm

• Five visual stimuli indicating different

reward probabilities: P=0,¼,½,¾,1

CS = 2 sec visual stimulus

US (probabilistic) = drops of juice

Fiorillo, Tobler & Schultz 2003

• At stimulus time: DA represents mean expected reward

• Interesting: A ramp in activity up

to reward (highest for p=½)

• Hypothesis: DA ramp encodes uncertainty in reward

Dopamine: Uncertainty or TD error?

• No apparent reason for ramp• The ramp is predictable from

the stimulus• TD predicts away

predictable quantities

contradiction !

• Side issue: the ramp is like a constantly surprising reward -- it can’t influence action choice

At time of reward:• Prediction errors result

from uncertainty

• Crucially: Positive and negative errors cancel out

A closer look at FTS’s results:

p = 0.5

p = 0.75

• TD error δ(t) can be positive or negative

• Neuronal firing rate is only positive (negative values are coded relative to base firing rate)

But:• DA base firing rate is low

-> asymmetric encoding of δ(t)

A closer look at FTS’s results:

55%

270%

δ(t)

DA

x(1) x(2) …

r(t) δ(t)

V(1) V(20)

• Tapped delay line • Standard online TD learning• Fixed learning rate

• Negative δ(t) scaled by d=1/6 prior to PSTH

Modeling TD with asymmetric errors

Learning proceeds normally (without scaling) − Necessary to produce the right predictions− Can be biologically plausible

TD learning with asymmetric prediction errors replicates

the recorded data accurately.

Ramps result from asymmetrically coded prediction errors propagating back to stimulus

Artifact of summing PSTHs over nonstationary recent reward histories

Modeling TD with asymmetric errors

Analytically deriving the maximum error at the time of the reward we get:

=> the ramp is indeed highest for P=½

But:• DA Encodes nothing but temporal difference error!• Experimental test: Ramp as within or between trial

phenomenon?

DA: Uncertainty or Temporal Difference?

)1)(1( dppTT

Trace conditioning: A puzzle and its resolution

• Same (if not more) uncertainty, but… no DA ramping! (Fiorillo et al.; Morris, Arkadir, Nevet, Vaadia & Bergman)

• Resolution: lower learning rate in trace conditioning eliminates ramp

CS = short visual stimulus

Trace period

US (probabilistic) = drops of juice

ConclusionsPreserve the TD hypothesis of Dopamine:

− No explicit coding of uncertainty− Ramping explained by neural constraints− Explains the disappearance of the ramp in

trace conditioning

Important challenges to the TD hypothesis − Conditioned inhibition− Effects of timing