Learning associations while retaining specificity: Competing demands on network plasticity rules.

Learning associations while retaining specificity: Competing demands on network plasticity rules.

Paul Miller, Brandeis University

with Mark Bourjaily (Neuroscience Program)

Talk Synopsis

1) Introduction to task and plasticity mechanisms

2) Associativity and specificity without persistence

3) Formation of sequential working memory

Introduction: paired associates

Introduction: paired associates

Introduction: paired associates

Introduction: paired associates

The Associative Inference Task (phases 1,2)

Bunsey and Eichenbaum 1996 Nature 379 25-257

Introduction: paired associates

Associative inference task phases 1-2.

Introduction: stimulus-response pools

Introduction: stimulus-response pools

Introduction: stimulus-response pools

Biased choicearises from reward-based plasticity?

Pair-responsivepools form via correlations of stimuli?

Introduction: stimulus-response pools

Questions to address

How can all these connections form?

Can they be stable once they have formed?

What biological plasticity rules help/don’t help?

Does initial network “architecture” matter much?

Plasticity mechanisms: STDP

Long-term potentiation of inhibition, LTPi

Maffei et al, Nature 2006

I to E connection increases with I-spike, veto by E-spike

E-cells, random sparseconnections

I-cells,all-to-all

2) Associativity and specificity without persistencea) Structured inputs

2) Associativity and specificity without persistenceProtocol, 2 secs of separate inputs A : B : A+B

Initial responses of selected cells in each pool2) Associativity and specificity without persistenceInitial responses

2) Associativity and specificity without persistenceCell responses after 200 trials: enhanced selectivity

A

B

AB

Post synaptic pool

Pre

syna

ptic

poo

l

Post synaptic pool

Pre

syna

ptic

poo

l

STDP alone destroys AB specificity (200 trials with homeostasis)

Inputs are all-to-all, randomized, no structure.AB selectivity arises with inhibitory plasticity alone

Summary of Part 2 (no persistence)

STDP tends to over-associateLTPi enhances specificity of neural activity

Inhibitory plasticity in an unstructured network can lead to specific stimulus-pair responses

Part 3 (formation of sequential working memory)

E-to-E plasticity certainly needed for persistence(cf GQ Bi’s work)

Best results with small-world connectivity:Mostly local E-E connections< 1% all-to-all E-E connections

STDP can generate “bumps” of memory activity.

3) Formation of sequential working memoryBump and ring attractor schematic

3) Formation of sequential working memoryBump and ring attractor schematic

3) Formation of sequential working memoryBump and ring attractor schematic

3) Formation of sequential working memoryBump and ring attractor schematic

3) Formation of sequential working memoryBump and ring attractor schematic

3) Formation of sequential working memoryBump and ring attractor schematic

3) Formation of sequential working memoryBump and ring attractor schematic

3) Formation of sequential working memory:Initial network response to X then Y

I-cells

Y

X

B

A

cue 2cue 1 Time (sec)

3) Formation of sequential working memory:Evolved network response to X then Y

Icells

Y

X

B

A

cue 2cue 1

cue 1 cue 2

3) Formation of sequential working memory:Evolved network response to A then Y

Icells

Y

X

B

A

3) Formation of sequential working memory:Evolved network response to A then B

Icells

Y

X

B

A

cue 2cue 1

Icells

Y

X

B

A

cue 1 cue 2

3) Formation of sequential working memory:Evolved network response to X then B

A-YA-BX-BX-Y

Stimulus-pair-dependent activity profiles (average firing rates between 1.5-2s after 2nd cue)

Some cells most responsive to X-B …

Some cells most responsive to X-B …

… only a few most responsive to A-B

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

M. Warden and E.K. Miller, Cereb Cortex 2007

M. Warden &E.K. Miller, Cereb Cortex 2007

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

3) Formation of sequential working memory:What about order of stimuli?

Icells

C

B

A

3) Formation of sequential working memory: orderEvolved network response to A then B

cue 1 cue 2

3) Formation of sequential working memory: orderEvolved network response to B then A

cue 1 cue 2

Icells

C

B

A

3) Formation of sequential working memory: orderEvolved network response to C then A

cue 1 cue 2

Icells

C

B

A

3) Formation of sequential working memory: orderEvolved network response to C then B

cue 1 cue 2

Icells

C

B

A

A-BB-A

C-AC-B

Future Directions Small-world topology without seeding?Generalize to more stimuli => higher-D topologyHow essential is each form of plasticity?

Are other stages of task with distinct roles of HC vs cortex explained by random vs small-world topologies?

AcknowledgmentsBrandeis Neuroscience Program, Swartz Foundation

F

Y

F

Y

F

Y

Y

F