Attractor Models of Cortical Computations (Working) Memory - - PowerPoint PPT Presentation

Attractor Models

• f

Cortical Computations

Master of Integrative Biology Neuroscience – UE 5BN04

2018-19 Gianluigi Mongillo

Center for Neurophysics, Physiology and Pathology Paris Descartes University – CNRS UMR 8119

(Working) Memory

Learning/Memory

Learning – process by which relatively permanent changes occur in behavioral potential as a result of experience. Memory – relatively permanent record of the experience that under- lies learning. Learning – acquiring new knowledge from experience Memory – retention of the acquired knowledge over time Internal representations

modification stabilization

neurons synapses

Models of Behavior

How do organisms select appropriate behavior?

stimulus response

reactive

“In mammals even as low as the rat it has turned out to be impossible to describe behavior as an interaction directly between sensory and motor processes” (Hebb, 1949)

Models of Behavior

How do organisms select appropriate behavior?

stimulus response

reactive

“In mammals even as low as the rat it has turned out to be impossible to describe behavior as an interaction directly between sensory and motor processes” (Hebb, 1949)

stimulus response internal representations

cognitive

The Cell Assembly Hypothesis

“Let us assume that the persistence or repetition of a reverberatory activity (or 'trace') tends to induce lasting cellular changes that add to its stability...”

• Perceptive experience activates sub-populations in neuronal assemblies,

by increasing/decreasing firing rates.

• Pairs of activated cells potentiate synapses between them, while synapses

from activated to non-activated cells are depressed.

• The resulting cell assembly is able to sustain a pattern of activity similar to

the perceptive one in absence of the eliciting stimulus. (Hebb, 1949)

The Hebb Framework

sensory motor association areas

The Hebb Framework

sensory motor association areas

The Hebb Framework

sensory motor association areas

reverberatory activity

The Hebb Framework

sensory motor association areas

reverberatory activity

The Hebb Framework

sensory motor association areas

reverberatory activity

Working Memory (WM)

cue delay period correct error

Delayed-response paradigm

test

Object Memory WM refers to the mechanism(s) underlying the maintenance of task-relevant information while performing the task. Items in WM are available in a special status, which makes them able to drive/control behavior (active maintenance).

+ + + + + + + + + + + +

correct saccade

cue delay period test

Spatial Memory

Working Memory (WM) Delayed-response paradigm

WM refers to the mechanism(s) underlying the maintenance of task-relevant information while performing the task. Items in WM are available in a special status, which makes them able to drive/control behavior (active maintenance).

DMS task and Delay Activity

(adapted from Meyer et al., 2007)

cue delay period correct error test cue delay period

cue delay period correct error test cue delay period

DMS task and Delay Activity

(adapted from Meyer et al., 2007)

cue delay period cue delay period correct error test

DMS task and Delay Activity

(adapted from Meyer et al., 2007)

(adapted from Funahashi et al., 1989)

ODR task and Delay Activity

Different Delay's Durations

(adapted from Funahashi et al., 1989)

Delay Activity During Error Trials

(adapted from Funahashi et al., 1989)

Mechanistic Accounts

Active maintenance as a result of the collective network dynamics Active maintenance as a result changes in single-cell excitability Active maintenance as a result of short-term modifications

• f synaptic efficacies

external input r e c u r r e n t c

• n

n e c t i v i t y neuronal population

Hebbian assembly

(Hebb, 1949; Amit, 1995)

activity time time activity

spontaneous activity memory activity

'active maintenance' through reverberatory activity

The Cell Assembly Hypothesis

cue period delay period

Persistent Activity and Network Multi-Stability

+ + + + + + + +

cue period delay period

Persistent Activity and Network Multi-Stability

+ + + + + + + + + + + + + + + +

A Toy Model

external input r e c u r r e n t c

• n

n e c t i v i t y neuronal population

Hebbian assembly

Si(t)=0,1

Si(t)=

0 otherwise 1 if I i(t)≥1 I i(t)=μext+ σext⋅ηi(t)+∑ j Jij S j(t−1)

all-to-all connected: Jij=J/N

A Toy Model

external input r e c u r r e n t c

• n

n e c t i v i t y neuronal population

Hebbian assembly

Si(t)=0,1

Si(t)=

0 otherwise 1 if I i(t)≥1 I i(t)=μext+ σext⋅ηi(t)+∑ j Jij S j(t−1)

all-to-all connected: Jij=J/N

ν(t)= 1 N ∑i Si(t) Φ(ν)= 1

√2π∫1−μext−J ν

σext ∞

dx e

−x

2/2

ν(t+ 1)=Φ[ν(t)]

with

Bi-Stability through Positive Feedback

ν(t) ν(t+ 1)

0.5 1.0 0.5 1.0

ν(t)

0.5 1.0

ν(t+ 1)

0.5 1.0

Bi-Stability through Positive Feedback

ν(t)

0.5 1.0

ν(t+ 1)

0.5 1.0

Bi-Stability through Positive Feedback

ν(t)

0.5 1.0

ν(t+ 1)

0.5 1.0 0.1 0.2 0.1 0.1

Bi-Stability through Positive Feedback

Realistic Models with Spiking Neurons

(Amit & Brunel, 1997; Brunel, 2000)

Network Architecture Single-cell Dynamics

AB Model: Population Activities

(Brunel, 2000)

sel.

• ther sel.

non sel. inhib.

AB Model: Single-Cell Spiking Patterns

sel.

• ther sel.

non sel. inhib.

(Brunel, 2000)

(adapted from Lisman & Idiart, 1995)

Through Single-Cell Properties

Firing is sustained by increased membrane excitability which is refreshed through network oscillations ADP time course

Acetylcholine Serotonin

V (t)=V osc(t)+V ADP(t)−V inh(t)

(adapted from Lisman & Idiart, 1995)

Through Single-Cell Properties

Firing is sustained by increased membrane excitability which is refreshed through network oscillations ADP time course

Acetylcholine Serotonin

V (t)=V osc(t)+V ADP(t)−V inh(t)

(adapted from Lisman & Idiart, 1995)

Through Single-Cell Properties

Firing is sustained by increased membrane excitability which is refreshed through network oscillations ADP time course

Acetylcholine Serotonin

V (t)=V osc(t)+V ADP(t)−V inh(t)

(adapted from Lisman & Idiart, 1995)

Through Single-Cell Properties

Firing is sustained by increased membrane excitability which is refreshed through network oscillations ADP time course

Acetylcholine Serotonin

V (t)=V osc(t)+V ADP(t)−V inh(t)

(adapted from Lisman & Idiart, 1995)

Through Single-Cell Properties

Sternberg effect (1966)

Response time increase linearly with the nr. of items in memory

(adapted from Lisman & Idiart, 1995)

Through Single-Cell Properties

Sternberg effect (1966)

Response time increase linearly with the nr. of items in memory

Nested oscillations (slow/fast)

Observed both in cortex and hippocampus:

• segment information in time
• time compression (for associations)