brain plasticity and the stability of cognition studies in cognitive neuroscience jaap murre...
TRANSCRIPT
Brain Plasticity and the Stability of Cognition
Studies in Cognitive Neuroscience
Jaap Murre
University of Amsterdam
Overview
• Background to two of our models
• Principles of multi-level modeling
• How our models are related
• How we obtain our data
• Research infrastructure and knowledge management
Background to two of our models
• TraceLink model
• Selfrepairing neural networks as a framework for recovery from brain damage
TraceLink model
Connectionist model of memory loss and certain other memory disorders
TraceLink model: structure
System 1: Trace system
• Function: Substrate for bulk storage of memories, ‘association machine’
• Corresponds roughly to neocortex
System 2: Link system
• Function: Initial ‘scaffold’ for episodes
• Corresponds roughly to hippocampus and certain temporal and perhaps frontal areas
Location of the hippocampus
System 3: Modulatory system
• Function: Control of plasticity• Involves at least parts of the hippocampus,
amygdala, fornix, and certain nuclei in the basal forebrain and in the brain stem
Stages in episodic learning
Sleep-consolidation hypothesis
• Memories are reactivated during slow-wave sleep
• This leads to a strengthening of their cortical basis
• After many weeks, the memories become independent of the hippocampus
• Unverified hypothesis: “Without such consolidation, memories remain dependent on the hippocampus”
Selfrepairing neural networks
A framework for a theory of recovery from brain damage
Redundancy and repair
• Redundancy by itself does not guarantee survival
• Only a continuous repair strategy does
• Example: safeguarding a rare manuscript
Redundancy and repair example
• Lesion: Suppose there is a 50% loss rate
Redundancy and repair example
• Repair: At the end of each month new copies are made of surviving information
This process has a long life-time
• Monthly ‘lesion-repair’ continues for many months ...
• ... until all information is lost at the end of one unfortunate month
• Chances of this happening are very low
• The expected life-time of the manuscript in this example is over 80 years
Application
• Spontaneous recovery
• Guided recovery: rehabilitation from brain damage
Studies in cognitive neurosciene
Principles of multi-level modeling
From brain to behavior
• Cognitive neuroscience, formerly called ‘Brain and Behavior’
• Question: How to bridge the gap between these two exceedingly complex objects of study?
• Partial answer: Through the construction of models
• But at what level should we model?
The problem
• Even simple behavior involves dozens of neural processes and structures with hundreds of parameters in total
• We are therefore forced to abstract from neural details
• Abstractions are based on assumptions about their – characteristics – interdependence
Detail and abstraction
• Verify assumptions with more detailed models
• Unfortunately: these simulations are very time consuming
• Therefore: show that they possess the essential characteristics that are assumed
• Low-level models are mainly suitable for verifying predictions at the level for which they have been developed
Principles of multi-level modeling
• We should model at several levels of abstraction
• Models at consecutive levels should be coordinated
• This is achieved by referring to the same concepts, processes, and structures
• Multi-level modeling is akin to having road maps at different levels of resolution
Multi-level modeling in cognitive neuroscienceLevel Brain Behavior1. Mathemati-cal
AbstractedNeuralSystems
Quantitative
2. High-levelConnectionist
NeuralSystems
Qualitative
3. Low-levelConnectionist
Details of oneor two systems
UnderlyingPrinciples
Level 1. Mathematical models
• Abstraction and generalization of TraceLink model with point process based models
• Investigation of possible neural basis of the REM model
Level 2. High-level computational models• TraceLink model
• Selfrepair model
• Hemineglect model
Level 3. Low-level computational models• Model of neural linking in the cerebral
cortex
• Hippocampus model
• Parahippocampus model
• Model of somato-sensory cortex
Illustration of different levels of modeling in our group
TraceLink as a starting point (level 2 model)
• Direct applications– Retrograde amnesia (loss of existing memories)
• Shape of the Ribot gradient (loss of recent memories)
• Strongly versus weakly encoded patterns
– Semantic dementia (loss of what things mean)• Inverse Ribot gradient (preservation of recent memories)
Extensions of TraceLink (level 2)
• Schizophrenia– Memory impairment is central in the ‘core
profile’ of schizophrenia
• Categorization– How and when should new categories be
formed
Detailing TraceLink (level 3)
• Trace system– Model of the formation of synfire chains: long-
range connections via a chain of neurons
• Link system– Hippocampal model– Parahippocampal model
• Modulatory system– Novelty-dependent plasticity
Example of a level 3 model
Synfire chain model
Formation of long-range connections in the cortex• If two remote brain sites A and B must
communicate via intermediary neurons, how is a communication path set up?
• Can such a path develop with normal learning?
Based on the work of Abeles: so called synfire chains
• Reliable transmission
• Increasing biological evidence
• The development of synfire chains, however, has not been simulated in a satisfactory manner
...Group 1 Group 2 Group 3
A B
Simulations
• We used a more biologically realistic model neuron (McGregor neuron)
• Self-organization of cortical chains was observed
Main characteristics of the development of synfire chains
• Chains develop with repeated stimulation of one or more groups
• A chain grows out of a stimulated group
• Early parts of a chain stabilize before late groups
Example of level 1 model
Point process model of learning, forgetting, and retrograde amnesia
(loss of existing memories)
Abstracting TraceLink (level 1)
• Model formulated within the mathematical framework of point processes
• Generalizes TraceLink’s two-store approach to multiple ‘stores’– trace system– link system– working memory, short-term memory, etc.
• A store corresponds to a neural process or structure
Learning and forgetting as a stochastic process• A recall cue (e.g., a face) may access
different aspects of a stored memory
• If a point is found in the neural cue area, the correct response (e.g., the name) can be given
LearningForgettingSuccessfulRecallUnsuccessfulRecall
Some aspects of the point process model• Model of learning and forgetting
• Clear relationship between recognition (d'), recall (p), and savings (Ebbinghaus’ Q)
• Multi-trial learning and multi-trial savings
• Massed versus spaced effects
• Applied to retrograde amnesia (hippocampus is store 1, which is lesioned)
• Applied to many learning and forgetting data
Hellyer (1962). Recall as a function of 1, 2, 4 and 8 presentations
0
0.2
0.4
0.6
0.8
1
0 10 20 30
Time (s)
Re
ca
ll p
rob
ab
ility
Two-store model with saturation. Parameters are1= 7.4, a1= 0.53, 2= 0.26, a2= 0.31, rmax= 85; R2=.986
Retrograde amnesia (RA)
• RA is loss of existing memories
• In current RA tests, questions about remote time periods are often easier than of recent time periods
• This makes them largely useless for modeling
• Our model can offer a solution because it can cancel the variations in item difficulty
Albert et al. (1979), naming of famous faces
a.
0
0.5
1
70s 60s 50s 40s 30s
Controls (N=15)Korsakoff's (N=11)Series3Series4
Example of multi-level approach
The same concept at three different levels
Learning associations between aspects of an experience
• Level 1. Increase of intensity through induction of ‘points’ (PPM model)
• Level 2. Hebbian learning between neural groups or ‘nodes’ (TraceLink)
• Level 3. Development of long-range cortical synfire chains (synfire chain model)
Obtaining data to model
Obtaining data to model
• Literature search
• Collaboration– Semantic dementia model: Cambridge group at
Medical Research Council - Cognition and Brain Sciences Unit
– Schizophrenia model: Washington Group at the National Institute of Mental Health
– Selfrepair and rehabilitation: Dublin group at Trinity College
Obtaining data to model: quantitative neuroanatomy
• Relatively little is known about mesoscopic aspects of the brain
• In particular, we do not know how neurons are connected
• We infer this mesoscopic level through mathematical modeling
• These data are of particular relevance for models at levels 2 and 3
Obtaining data to model: retrograde amnesia (RA)
• No RA tests in Dutch. Therefore:– Official translation of British test– Public events test
• Novel aspect: using the internet to obtain data on long-term forgetting (Daily News Test)
Direct investigation of consolidation: sleep experiment
• Consolidation lies at the heart of the PIONIER projects
• Much circumstantial evidence for the existence of memory consolidation during sleep
• No direct evidence
• Therefore: investigate this ourselves
• Also: makes integration of our group with the neurosciences more of a reality
Research infrastructure and knowledge management
Infrastructure for research and knowledge management
• Simulation software
• Dissemination of results
• Preservation and exchange of knowledge within the group
Neurosimulation software developed by us: Walnut and Nutshell
• Aimed at users in cognitive neuroscience
• Greatly shortens development cycle of new models
• Useful to both naïve and expert users
• Exchange of paradigms and simulations across the internet via NNML
• Scriptable in VBScript, Python, etc.
Dissemination of results
• How to publish or obtain models?
• Geppetto project: ‘Bring models to life’
• Database of – models– neurosimulators (modeling software)– data– researchers and laboratories
Dissemination of results (cont’d)
• Presentation of the PIONIER group’s activities
• neuromod.org (neuromod.uva.nl): research
• memory.uva.nl: general audience
Preservation and exchange of knowledge within the group
• Intranet for within-group cooperation and exchange
• Database management (with backups etc.)
• Documentation of procedures
• Version control system (great ‘Undo’)
• Issue and task management (e.g., bugs)
• HowTo texts
Concluding remarks
Modeling in a multi-discipline
• Our models incorporate data from:– Neuroanatomy and neurophysiology– Neurology and neuropsychology– Experimental psychology
• The ultimate aim is to integrate these various sources of data into a single framework that is implemented as a series of coordinated models
Steps towards the goal
• In the following two hours, we will present some of our progress made towards that goal