PPT - Biointelligence Lab.

Download Report

Transcript PPT - Biointelligence Lab.

9 Distributed Population Codes in
Sensory and Memory Representations of
the Neocortex
Matthias Munk
Summarized by Eun Seok Lee
Biointelligence Laboratory, Seoul National University
http://bi.snu.ac.kr/
Summary




Not single-unit level, but neuronal coding and integrative
processes in neocortex are presented.
Distributed representations are inevitable.
2nd level organization – mesoscopic level of neuronal
signal transformation and processes is needed for
complex and adaptive functions of memory
Cortical network operations at the mesoscopic level: 3
kind of activity patterns are presented.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
2
Introduction

Distributed representations are inevitable from
 Large neuronal circuits to a detailed and adaptive analysis of
complex information



Optimized dynamical structure and operation of multiple
parallel neuronal processes
The concept – information coding is based on tuning
functions of many individual neurons thought to express
their stimulus specificity in an independent way.
2nd level organization – spatial relation of neurons,
convergence of neuronal signals from different
modalities into “higher” areas of executive functions or
the complex memory formations
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
3
Introduction



Collection of neuronal signals be used to reconstruct
population codes like ‘population vector’ analysis in
motor, sensory, memory areas of cortex
Brain processes rapidly changing relations among
content and context for which more dynamic neuronal
representations are required
We explore mesoscopic range of cortical signals that
might contribute to distributed codes in the cortex with
particular emphasis on the malleability of representations
as the underlying mechanism of learning processes.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
4
Why is neuronal coding a difficult issue?




Central problem in neuroscience: How the brain or neocortex codes
information and how the signals are used by neuronal processes for the
control of behavior
“self-referencing system” “ongoing self-maintaining system” – so treating
brain as an input-output system can have only limited success.
Many studies in neuronal procedure and functional mechanism lacks
“comprehensive understanding of the embedding of these many partial
functions into ongoing processes or, into a “brain’s life” that all prior
experience and memories may influence processing.
Neuronal coding, the embedding of specific functions requires that
representations which use a particular neuronal code are sensitive to the
context, but at the same time allow for a sufficient degree of invariance to
make perceptual processes reliable and therefore support recognition.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
5
Why is neuronal coding a difficult issue?

A simple distributed code would not work and the system
needs to organize distributed activity in such a way that
different processes can run simultaneously without
interference in the same network – how ensembles of neurons
can self-organize for this purpose
 Neuronal coding has to involve constant dynamical
restructuring of activity which has to incorporate contextual
information and therefore has to include many distributed
circuits if not the entire brain without loosing the specificity
of concurrent processing
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
6
What properties does a neuronal code need
to have?


Code – an instruction or rule with which a signal that contains
info is transformed into another signal that can serve as a
message or command which is transmitted to a receiver or
target system.
Neuronal code can only be deciphered if, in addition to the
transmitted neuronal activity patterns that have been
generated by applying the rules of a particular code, the
neuronal readout mechanism has been identified and used to
interpret the transmitted signals.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
7
Established and proposed codes





Code for occurrence of sensory stimuli has been considered as most
fundamental historically.
Not single neurons. Codes have been proposed which make use of activity
coordinated among distributed populations of neurons.
Neuronal processing itself is in general highly distributed and may cause
high energy demand due to the need for coordination among remote
neuron populations.
How the brain codes information – distributed spatial patterns of activity
seem to reflect a coding strategy
As spatial and temporal resolution of functional imaging techniques
improves, more detailed information about the macroscopic and
mesoscopic aspects of distributed representations not with micro-electrode
experiments
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
8
Cortical network operations at the
mesoscopic level: 3 activity patterns





Promising candidate signals for serving neuronal
computations
Depend on intact cortical connectivity, and its own
dynamical properties
1. Cortical avalanches: represent waves of transmitted
synchronous activity. Most recent descovered patterns.
2. Spatio-temporal spike patterns as a correlate of synfire
chains: the concept of how cortical connectivity might
determine brain activity
3. Oscillations: the oldest form of brain activity.
Important constituent of modern concepts of brain
function (C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
9
Cortical network operations at the
mesoscopic level - Avalanches




Definition: Propagating synchronized signals of large
neuronal populations
Good evidence for successful transmission processes at
the population level
Depending on it, neuronal network can be in a “critical”
state: the diversity of processes with respect to their size
and duration is scale-free.
Patterned neuronal activity is not confined to the
stimulus or behavior-related processes, but relevant for
intrinsic and ongoing processes: in human brains, entire
territories seem to be devoted to processing intrinsic
signals, not to the well-known sensory domains.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
10
Cortical network operations at the
mesoscopic level - Avalanches
Fig. 9.1.
(a) Avalanches of different length and size
(b) Size distribution of avalanches
(c) Branching pattern of a self-organizing critical process
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
11
Cortical network operations at the
mesoscopic level - Avalanches
Fig. 9.2.
Intrinsic and extrinsic
cortical systems
related to natural
scene stimulation.
Extrinsic: driven by
sensory stimulation in
different modalities
Intrinsic: territories
within the gaps of the
extrinsic system
identified by highly
correlated activity
among each other
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
12
Cortical network operations at the
mesoscopic level – Synfire chains



Synchronous transmission in a stable fashion
Feature: Ubiquitous divergence and convergence of
cortico-cortical connectivity
Neurophysiological correlate of synfire chains: precise
“spatiotemporal spike patterns” occurred in frontal areas
Fig. 9.3.
(a) Network
supporting synfire
activity
(b) Spatiotemporal
patterns recorded
from frontal cortex
(cell 12  cell 2
 cell 13)
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
13
Cortical network operations at the
mesoscopic level - Oscillations


Cortical oscillations be useful for understanding how
large populations of cortical neurons may be coordinated
so that they can directly code information or support
complex neuronal operations
State dependence found in EEG study: rhythms at other
frequency ranges are to some extent state dependent, but
may participate in or support neuronal coding of
information.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
14
Cortical network operations at the
mesoscopic level - Oscillations
Fig.9.4. Oscillations: Lateral aspect of a macaque brain with recording positions
(upper left) and averaged autocorrelation (gray) and cross-correlation (white)
functions with shift predictors (lower right) computed from transcortical field
potentials recorded simultaneously in four cortical areas.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
15
Population codes and distributed
representations


Representations for complex visual objects are distributed over large
populations of neurons in inferotemporal cortex.
Tuning of individual cells is broad, but the acuity of the population of cells
is close to identifying individual faces.
Fig. 9.5. Distributed and sparse
representation. (a) Neuronal
population data after
multidimensional scaling (b)
Parameters used to generate a
physical model of the faces (c)
Population vectors indicating
the match between model and
neuronal data.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
16
Visual representations during short-term
memory: units versus populations




Memorizing mechanism  the stabilization of activity patterns
Not by a single dynamical principle like sustained spike firing, but
integration of many different dynamical aspects in an adaptive fashion
provide sufficient stability of representations
Multiple intricate mechanisms have to cooperate to achieve adaptive
memory performance
Populations of neurons may be a useful level of organization for a better
understanding of memory mechanisms
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
Fig. 9.6. Stimulus
selectivity and
distributed activity
patterns in lateral
prefrontal cortex of
monkeys performing
a simple visual
memory task.
17
Visual representations during short-term
memory: units versus populations


Comparing time course of stable patterns/ firing rates: different groups are
co-activated during different epochs with increased frequency of stable
patterns.  different groups of neurons cooperate in order to constitute the
distributed activity patterns at the population level.
Multiple cortical and subcortical areas participate in al complex cognitive
functions.
Fig. 9. 7. Strength of
ensemble selectivity
and simultaneous
modulation of firing
rates.
(C) 2009, SNU Biointelligence Lab, http://bi.snu.ac.kr/
18