Mode shifting between storage and recall based on novelty d - PowerPoint Presentation

Slide1

Mode shifting between storage and recall based on novelty detection in oscillating hippocampal circuits

M.

Meeter

J. M. J.

Murre

L. M.

Talamini

Date of Presentation: 05/09/2012Slide2

IntroductionSlide3

Role of Acetylcholine in Mode Shifting

Hippocampal

novelty detection may regulate levels of acetylcholine

Shifts

hippocampal

dynamics between encoding or retrieval

Information with high novelty content would induce a learning state

Input similar to already stored patterns would induce a retrieval state

Little learning takes place during retrieval to protect existing patterns from modification.Slide4

Role of Acetylcholine in Mode Shifting

Hippocampal

novelty detection may regulate levels of acetylcholine

Shifts

hippocampal

dynamics between encoding or retrieval

Retrieval mode: low

ACh

levels

Learning mode: high

ACh

levelsSlide5

Role of Acetylcholine in Mode Shifting

Hippocampal

novelty detection may regulate levels of acetylcholine

Shifts

hippocampal

dynamics between encoding or retrieval

Information with high novelty content would induce a learning state

Input similar to already stored patterns would induce a retrieval state

Little learning takes place during retrieval to protect existing patterns from modification.Slide6

Learning StateSlide7

Role of Acetylcholine in Mode Shifting

Hippocampal

novelty detection may regulate levels of acetylcholine

Shifts

hippocampal

dynamics between encoding or retrieval

Information with high novelty content would induce a learning state

Input similar to already stored patterns would induce a retrieval state

Little learning takes place during retrieval to protect existing patterns from modification.Slide8

Retrieval StateSlide9

Opposition to ACh’s Role in Mode Shifting

ACh

has a slow and sustained influence on

hippocampal

activity

Depolarization develops a few seconds after

ACh

release

Lasts 10 s or more

Dynamics are too slow to underlie mode shifting.

Opposers

support a fast mode shift

On the order

ot 10s or 100s of millisecondsSlide10

Opposition of ACh’s Role in Mode Shifting

ACh

has a slow and sustained influence on

hippocampal

activity

Depolarization develops a few seconds after

ACh

release

Lasts 10 s or more

Dynamics are too slow to underlie mode shifting.

Opposers

support a fast mode shift

On the order of 10s or 100s of millisecondsSlide11

Opposition to the Opposition of ACh’s Role in Mode Shifting

Time scale at which natural learning and retrieval take place is slow

Rats take minutes to familiarize themselves to new environments

Fear conditioning takes several seconds to take (in rats)

In humans, long-term recall deteriorates when study times are less than 2 sSlide12

The SystemSlide13

Structural and functional properties of the subregions and connections

Feedback and

feedforward

inhibition

Oscillatory population dynamics

Theta (4-10 Hz) and Gamma (20-40 Hz)

Features

Integrate-and-fire nodes (Sodium, potassium, chloride, and leak currents

Hebbian

learning (LTP) and negative

Hebbian

learning (LTD)Slide14
Slide15

Entorhinal CortexMain cortical input structure of the hippocampus

Considered as one single input layer that propagates the same information to the DG, CA3, and CA1Slide16
Slide17

Dentate Gyrus

Receives the majority of the input from the EC

Sparseness of activation

High number of granule cells but very few fire at a given moment

Divergent input

Orthogonalizes

input

Orthogonalization

Patterns that are correlated in a given layer (i.e. EC) generate uncorrelated representations in the projection field (i.e. DG)

No

Hebbian

plasticity between the DG and CA3

Feedforward inhibition to CA3Slide18
Slide19

CA3

Involved in either:

Autoassociative

learning and pattern completion

Heteroassociative

learning and sequence recall

Both cases are inferred from extensive recurrent connections among CA3

Difference in time scale:

Autoassociation

: learning over short intervals

Heteroassociation

: learning over longer intervalsSlide20
Slide21

CA1Receives a direct projection from the EC (

Yeckel

and Berger, 1990)

Indirect projection from tri-synaptic loop

Proposed to be a translator between the code of the CA3 and the cortical code

Model associates the CA3 pattern with the EC patternSlide22
Slide23

Medial Septal Nuclei

Fibers from the CA1 and CA3 target

GABAergic

septal

projection neurons and

ACh

neurons

Hippocampal

activity inhibits the septum

Modulates the hippocampus using

ACh

Nonspecific targetSeems to affect the entire hippocampusSlide24

Hippocampus

Summary

EC:

- Major input

DG:

-

Orthogonalizer

CA3:

- Storage

CA1:

- Translator

Septum:

-

ACh

modulationSlide25
Slide26

Acetylcholine as a Novelty SignalExperiments

During exploration of a new environment,

ACh

is increased relative to baseline

ACh

levels decrease during consecutive explorations of the same test

enviornmentSlide27

Effects of Acetylcholine

Dampens transmission between CA3 and CA1

Slow,

subthreshold

depolarization of

hippocampal

principal neurons lasting several seconds

Enhancement of LTP at CA3, CA1, and DG

Reduction of adaptation in CA3, CA1, and DG

Suppressed inhibition of DG and pyramidal cells (

supression

of basket cells)Slide28

Effects of Acetylcholine

Learning mode

High

ACh

CA3-CA1 transmission dampened

Activity in CA1 is dominated by EC

Allows CA3-CA1 connection to store the association between the EC pattern and CA3 pattern

Retrieval mode

Low

ACh

CA1 relays the reinstated CA3 pattern to the EC and other output structuresSlide29

ResultsSlide30

StorageSlide31

“Correct CA1 Nodes”

Correct CA1 Nodes: number of firing CA1 nodes that receive a one-to-one connection from an EC node

Incorrect CA1 Nodes: number of firing CA1 nodes that are not connected to an EC node

Missing CA1 nodes: number of CA1 nodes that receive a one-to-one connection from an EC node that does not fireSlide32

Retrieval

R

etrieval mode (

ACh

= 0.1)Slide33

Pattern Completion

After storing one pattern, a variable number of EC nodes associated with the pattern are deactivated

Test in retrieval mode (

ACh

= 0.1) and learning mode (

ACh

= 0.75)

Pattern completion measured as the maximum proportion of correct CA1 nodes that were simultaneously active during the theta cycleSlide34

Pattern Completion

CA1 activity is strongly correlated with DG activity

Pattern completion occurs in DG

Number of correct nodes increases

Number of incorrect nodes increases

ACh

depolarizes all cells making activation easierSlide35

Pattern Completion

With small cue-sizes, more of the pattern is completed in learning mode than in retrieval mode but comes at a price of a compromised integrity of retrieval

Hypothesis: During effortful retrieval, input cues do not lead to an instatement of a stored pattern, so the hippocampus shifts to learning mode

Consistent with the “retrieval practice effect”

Implies that effortful and successful retrieval constitutes a power learning method

Effort retrieval has a stronger effect on retention than fast, easy recallSlide36

Novelty Detection and Dentate Gyrus

After acquisition of one pattern, the network was cued with patterns that were either the same as the stored pattern (old), completely different (new), or consisted of a variable ratio between old and new EC nodes

DG same: Same DG nodes activated as during the stored pattern

DG diff: Different DG nodes activated than with stored pattern

Correct CA1 nodesSlide37

Novelty Detection and Dentate Gyrus

DG and CA1 liked old stuff, not new stuff

Little overlap under areas of same and diff

All types of input elicit strong activity in DG and CA1

Larger range of mixed input activates mixture of old and new DG nodesSlide38

Effects of ACh on Pattern Storage

New pattern was presented during one theta cycle with different levels of

ACh

modulation

Acquisition was evaluated by presenting the pattern during one theta cycle with

ACh

modulation set at 0.1

Maximum number of correct CA1 nodes simultaneously active during retrieval phase was used to assess

ACh

effects on learning performance

If

ACh

was too high during learning phase, then learning was inhibited

If

ACh

is too high, then EC alone can cause CA1 to fire before CA3 can stimulate CA1

CA1 cells that do not receive EC input will form stronger connections with their CA3 input, which does not reflect EC firing for diminished retrieval performanceSlide39

Effects of ACh on Novelty Detection

Stored one pattern and presented either the “old” pattern or a randomly selected new pattern while varying

ACh

modulation

Look at difference in activity from the old pattern versus the new patternSlide40

Effects of ACh on Novelty Detection

Novelty detection decreases as

ACh

increases

Difference in activity is first seen in DGSlide41

Insights from the Model

Prominent role of DG in novelty detection

Reasoning for the existence of separate learning and retrieval modes

Retrieval in learning mode is unreliable through the activation of features that were no part of the original memory

Separation of learning and retrieval modes enhances accurate retrieval

Hippocampus incorporates a low band filter to ensure that

ACh

modulation fluctuates with novelty and not with theta or gamma rhythmsSlide42

Slow Shift vs. Fast Shift

Slow Shift

Mediated by

ACh

Mode shifting induced by novelty of input

Most learning takes place when a pattern is new

Fast Shift

Mediated by theta modulation where LTP and LTD dominate in different phases of the cycle

Occurs automatically due to differing learning and retrieval dynamics on different phases of theta

Old and new patterns are continuously learned, unlearned, and relearnedSlide43

Predictions of the Model

Learning should occur more quickly in experimentally induced high

ACh

states versus low

ACh

states

Novel configurations lead to increased activation of cholinergic cells leading to a surge in

hippocampal

ACh

release

Novel configurations produce enhanced synaptic plasticity in the hippocampus

New patterns should elicit little activity in the hippocampus in the absence of AChModel suggests that balance of strength between direct perforant

path and trisynaptic input to CA1 is essential to pattern encodingSlide44

Other Possible Novelty SignalsNovelty signal over the CA3 through intermediaries of lateral septum, and

raphe

nuclei, reticular formation

Could influence levels of arousal

Novelty signal via the EC to the ventral striatum

Could influence chain of events that facilitate a change of behavioral strategySlide45

Uses of the Model

Explore the significance of different parallel inputs (input from different layers of the EC) to the hippocampus for memory processing

Explore how

autoassociation

and

heteroassociation

may be implemented in circuitry

How suppression of familiar objects in

parahippocampal

cortex affects configuration novelty detection in hippocampus

How

hippocampal

subdivisions differentially contribute to neuropsychological constructs such as recall, recognition, and familiarity processingSlide46

Problems

“Therefore, at present there is no direct evidence for a

hippocampal

influence on

ACh

release during behavioral learning.” !!!!!!!!!

“That is, the

GABAergic

cells would

disinhibit

ACh

cells in response to hippocampo-septal stimulation.”Opposite of what the model says it should do“However, the inhibitory hippocamp-medioseptal

projection directly onto ACh neurons may be sufficient to regulate activity of ACh neurons during mode-shifting.”Slide47

Extra StuffSlide48
Slide49
Slide50
Slide51