Principles of Neural Data Processing in the Brain ?
Heinz Horner
Institute für theoretische Physik Ruprecht-Karls-Universität Heidelberg
horner@tphys.uni-heidelberg.de
Different levels of understanding
Consciousness, emotions free will, behaviour
...
...
Neocortex, midbrain, cerebellum
Cortical areas, nuclei, long distance wiring Cortical layers, columns short distance wiring Neurones, (glia cells),
communication among neurones Dendrite, soma,
axon, synapse
Membrane, ionic channels, ionic pumps, receptors, neurotransmitters
...
Anatomy of the human brain, cortical Areas
Corpus collosum Connection between
hemispheres Cingulate gyrus
Emotional, cognitive and motor tasks
Thalamus
Relay station to the cerebral cortex
Hippocampus Long term memory Maps for navigation Fornix
Formation of memory
Mammillary body Formation of
memory Amygdala
Emotions fear, reward Hypothalamus
Temperature sleep ...
Olfactory bulb Smell
Stria terminalis Output from
amygdala
Brainstem Prefrontal
cortex Planing
Motor cortex Motion
Somatosensory cortex Sense of touch
Parietal lobe Integration of
sensory information
Visual cortex Vision Angular gyrus
Verbal association
Cerebellum Coordination
of motions Wernicke‘s area Auditory cortex
Hearing Temporal lobe
Organisation of sensory input Memory Broca‘s area
Speech
Limbic system
Cerebral cortex
Cortical layers
Grey matter
Cell bodies:
Neurones, Glia cells Neuropil
White matter
Myelinated axons (long distance connections)
5 mm
Corpus collosum Connection between
hemispheres Cingulate gyrus
Emotional, cognitive and motor tasks
Thalamus
Relay station to the cerebral cortex Fornix
Formation of memory
Mammillary body Hypothalamus
Temperature sleep ...
Olfactory bulb Smell
Stria terminalis Output from
amygdala
Brainstem Respiration Blood circulation Prefrontal
cortex Planing
Motor cortex Motion
Somatosensory cortex Sense of touch
Parietal lobe Integration of
sensory information
Visual cortex Vision Angular gyrus
Verbal association
Cerebellum Coordination
of motions Wernicke‘s area Spoken language Auditory cortex
Hearing Temporal lobe
Organisation of sensory input Memory Personality Broca‘s area
Speech
Long ranged connections
Visual pathway
Cortical layers
Pyramidal neurones (c) excitatory
long ranged connections
Spiny stellate cells (d) excitatory input from thalamus
Basket cells (e) inhibitory project to soma of
pyramidal neurones
Marinotti cells (f) inhibitory project into layer I
Associative fibres (a)
from other parts of the cortex
Specific fibres (b)
from thalamus
Axons from pyramidal cells
I II
III
IV
V
VI
Cell types:
Afferent fibres:
Efferent fibres:
Synapse
inhibitory
Dendrite Synapse
excitatory
1 µ
Grey matter
Coarse wiring:
specific,
genetically predetermined
Wiring in detail:
determined by chance and learning
Cortical wiring
Thickness of the cortex 3 - 4 mm Area of the cortex 0.5 qm Neurons 10 - 10 Neurons per mm 10 Synapses per neuron 10 Synapses 10 - 10 Length of dendrites per neuron 10 mm Length of dendrites per mm 400 m Length of axons per mm 3000 m Time scale 10 - 100 msec Computing power > 1 PetaFLOPS Storage capacity > 100 Terabyte Energy consumption 20 W
3
3 3
10 11
5 4
14 15
Some numbers
?
McCulloch-Pitts neuron, perceptron
Presynaptic neurons 1 · · · N
Activity, firing rate fi
Synaptic efficacy Wi
f f
1f
if
NPotential U = X
i
Wi fi
Threshold #
Firing rate f = ⇥(U #)
S Dendrite:
incoming signals, membrane potential
Soma:
threshold
Axon:
outgoing signals, spikes
Decision plane
X
i
Wifi = #
f2
f1 S W~
o +
N = 2
Perceptron learning
Learning rule: change couplings W
i⇠ ± ⇠
iµf
2f
1S
W ~
+
o
W ~
µ⇠ ~
µHebb’s rule: (D.O.Hebb 1949, S. Freud, S. Exner 1895)
If an axon of cell A is close enough to cell B, such that B is repeatedly excited by A and firing, a change takes place such that the ability of cell A to stimulate cell B is increased.
⌘
µ= 1
Experiment: Bi, Poo (1998)
Learning patterns with desired output
K⇠
iµ⌘
µ= { 1, 0 }
Estimate of the capacity of the perceptron
How many (random) patterns with mean activity can be classified ?a Probability and .P⌘=0=a P⌘=1= 1 a
For random couplings the probabilities for an actual output are and .
Wi f ={1,0} Pf=1=a Pf=0= 1 a
This means . P⌘=1,f=0=P⌘=0,f=1=a(1 a)
For patterns a subset of patterns has to be ‘embedded‘ , i.e.
the couplings have to be determined such that
A Ax = 2a(1 a)A
This results in linear equations for the couplings . They can be solved for
N Wi 2a(1 a)A
A 2a(1N a)
X
i
Wi⇠iµ=# ± "
Estimate of the capacity of the perceptron
Shannon information
0 2 4 6 8 10
0.01 0.1
I
a 0.5
The total number of possible patterns with mean activity isa
Perceptron:
linear separable classification,
increased capacity and information content for sparse coding
N!
(aN)! ((1 a)N)! ⇠e{aln(a)+(1 a) ln(1 a)}N
N/2a(1 a)
Capacity with and .c(a= 12) = 2 c(a! 1)⇠ 2a1
I
N = 12n 1
1 a log2(1a) + 1a log2(11a)o c= AmaxN
c= AmaxN
Iterate presenting patterns modify couplings only if according to Hebb‘s rule
⇠~µ
fµ6=⌘µ
µW~ =± ⇠~µ
a⌧1
⌘µ= 1 ⇠iµ= 1
Perceptron learning
Learning:
learning regulated by novelty, reward, e.t.c reduced learning precision for sparse coding
W
W
⇠
aN1Modification for sparse coding : Learning only if and
Required accuracy for each learning step:
Divergent preprocessing Input layer
NOutput perceptrons Hidden layer
LExample:
Perceptron with random preprocessing:
The capacity is determines by the size of the hidden layer. It is able to handle non linear separable problems.
L
Divergent preprocessing:
increased capacity,
preprocessing random or determined by unsupervised learning, sparse coding in the hidden layer.
Liquid computing
Cortex:
Almost all connections are intracortical, for exampletotal number of connections
connections between hemispheres cortex cerebellum
optical nerve ear cortex 108
107 106 105
1014
HD-Stud.Days April 2013 16
Associative memory, attractor network
Associative recall:
Any part of the items memorized can trigger the recall of the complete item.
Unlike in computers there is no address under which items are stored.
The storage is distributed. A single item is shared by many neurons.
A given neuron can be part of many items stored.
Attractor network, Hopfield model:
Fully or partially connected network of McCulloch-Pitts neurons
Construct an “energy” such that the patterns to be stored form
local minima, surrounded by
basins of attraction (gradient dynamics)
Kazimierz 10-17.06.05 45
Different forms of long term memory
conscious unconscious
Kazimierz 10-17.06.05 46
Associative recall:
Any part of the items memorized can trigger the recall of the complete item
Unlike in computers there is no address under which items are stored The storage is distributed. A single item might reside in several
areas of the cortex
Attractor network, the Hopfield model
(Fully) connected network of McCulloch-Pitts neurons or
Construct an “energy” such that the patterns to be stored form
local minima, surrounded by
basins of attraction (gradient dynamics)
f(U) = Θ(U) sign(U)
The Hopfield model
Patterns to be memorized: random Hebb learning:
Firing rate: or
Membrane potential:
Temporal evolution: sequential updating
“Energy”, cost function:
Change of energy: for
for
ξiµ = ±1 Wij = N1
!A
µ=1
ξiµξjµ
fi =
!1
0
" ! 1
−1
"
Ui =
!
j
Wijfj =
!
µ
ξiµmµ mµ = N1
!
j
ξjµfj
fi(t + τ) = sign!
Ui(t)" E(t) = −12
!
ij
fi(t)Wijfj(t) = −N2
!
µ
mµ(t)2
fi(t)Ui(t) < 0 fi(t)Ui(t) > 0
∆iE(t) = −2 sign!
Ui(t)"
Ui(t) < 0
∆iE(t) = 0
Montag, 15. April 13
Attractor network, Hopfield model, sparse patterns
Signal to noise analysis:
assume the firing state is near the pattern :⇠i fi=⇠i m
potential Ui=X
1
Wijfi= ⇠i m + X
µ6=
⇠iµmµ Patterns to be memorized: with =
⇢1
⇠iµ 0 ⌦
⇠iµ↵
= a ⌧ 1 µ= 1..A
Hebb‘s learning: Wij = 1
aN
X
µ
⇠µi ⇠jµ
‘overlap‘ between actual firing state and pattern :fi ⇠iµ mµ= aN1 X
i
⇠iµfi
average over random patterns :⇠iµ hmµi=a m hm2µi hmµi2= 1
N m2
2(m ) =hUi2i hUii2 = aAN m2 hUii⇠ = ⇠i m + ¯U
The distribution is a Gaussian with width centered around P⇠(Ui) (m ) hUii⇠ P⇠(U) = 1
p e (U 2 2⇠m )2
Attractor network, Hopfield model, sparse patterns cont.
P⇠(U) = 1
p2⇡ (m)2 e
(U ⇠m)2 2 (m)2
# U
a P1(U) (1 a)P0(U)
m
Retrieval dynamics: (m)
@tf⇠= Z
#
dU P⇠(U) f⇠
Retrieval dynamics:
with constraint a f1 + (1 a)f0 = a
non trivial solution for information per synapse A < N 2 aln(a1)
I
N2 a!
!0
1
2 ln(2) ⇡ 0.72
Attractor network:
patterns are stored as fixed points of retrieval dynamics
reduced learning precision for sparse coding
Dynamic attractors, transients
Non symmetric couplings Retarded couplings
Fatigue after ongoing firing
Forward excitation and inhibition Special delay lines (Cerebellum)
Fixed points (associative memory) Limit cycles (rhythm generator) Transients (motion generator,
short time memory, liquid computing) Special delay lines (Cerebellum)
Wij 6= Wji Ui(t) = X
j
Wij(⌧)fj(t ⌧)
t
#i(t)
fi(t)
Dynamic attractors, transients
various tasks various mechanisms
Winner takes all mechanism
Task: select the neuron(s) with the highest input Excitatory neurons with input
hiInhibitory neuron
i
o
Synapses
Wio=W Woi= XF(t + ⌧) =X
i
fi(t + ⌧) =X
i
⇥(hi X fo(t) #) fo(t + ⌧) = ˜⇥(W F(t) #o)
F(fo) fo
F fo(F)
Winner takes all:
unsupervised learning retinotopic and other maps
vector quantization
can lead to oscillations (limit cycle)
Visual system:
mapping retina NGL V1
Topology preserving (retinotopic) maps
Somatosensory cortex Motorcortex
Auditory system:
tonotopic maps
Topology preserving (retinotopic) maps
Input space
Target space x
y
Topology preserving mapping of an input space onto a target space (Kohonen map)
x y
Initial state:
no structure in the couplings
Apply correlated signals at random
topology in input space is established by the patterns
Target space:
lateral short ranged excitatory couplings
establishes topology in target space
winner takes all mechanism
Hebb‘s learning for
W(x, y)
xo
⇠(x) = '(x xo)
W(x, y)
Topology preserving (retinotopic) maps
Topology preserving mapping of an input space onto a target space (Kohonen map)
x y
Input space
Target space x
y
Initial state:
no structure in the couplings
Apply correlated signals at random
topology in input space is established by the patterns
Target space:
lateral short ranged excitatory couplings
establishes topology in target space
winner takes all mechanism
Hebb‘s learning for
W(x, y)
xo
⇠(x) = '(x xo)
W(x, y)
Topology preserving maps:
unsupervised learning
with correlated input patterns
HD-Stud.Days April 2013 24
Experiment on awake monkeys:
three types of dynamic dot patterns ( ) are repeatedly presented. The spike trains of a neuron in the visual cortex are recorded. Each pattern is presented 50 times. The spike trains are plotted individually as function of time past
beginning of the presentation.
The histogram exhibits clear pattern dependent structures, whereas individual spike trains exhibit strong fluctuations.
This and similar experiments suggest coding by firing rates of subpopulations.
Single spike or firing rate coding?
Bair and Koch 11
j024 140
spk/s c=1.0c=0.5c=0.0
0 500 1000 1500 2000
Time (msec)
Figure 5:Temporal modulation disappears for highly coherent stimuli. The spike trains and PSTHs demonstrate that the stimulus-locked temporal modulation present for incoherent motion (c=0) and for partially coherent motion (c=0.5) was virtually absent during the sustained period of the response to coherent motion (c=1). This suggests that temporal dynamics of a higher order than those found in rigid translation are necessary to induce a specifi c and unique time course in the spike discharge pattern.
Bair and Koch 11
j024 140
spk/s c=1.0 c=0.5 c=0.0
0 500 1000 1500 2000
Time (msec)
Figure 5: Temporal modulation disappears for highly coherent stimuli. The spike trains and PSTHs demonstrate that the stimulus-locked temporal modulation present for incoherent motion (c = 0) and for partially coherent motion (c = 0 . 5) was virtually absent during the sustained period of the response to coherent motion (c = 1). This suggests that temporal dynamics of a higher order than those found in rigid translation are necessary to induce a specifi c and unique time course in the spike
100ms c = 0.0, 0.4, 1.0
Coding by firing rates of subpopulations:
structures on time scale ~ 20 ms mean firing rate < 100 Hz Exceptions? Auditory system?
Montag, 15. April 13
HD-Stud.Days April 2013 25
The role of inhibitory neurons, balanced networks:
Experiment:
Fluctuations of the membrane potential (visual cortex of a cat)Ui(t)1.5 The Neural Code 33
in vitro in vivo in vivo
20mV 100 ms
current injection current injection visual stimulation
Figure 1.17: Intracellular recordings from cat V1 neurons. The left panel is the response of a neuron in an in vitro slice preparation to constant current injection.
The center and right panels show recordings from neurons in vivo responding to either injected current (center), or a moving visual image (right). (Adapted from Holt et al, 1996.)
be captured by the Poisson model of spike generation, the spike generating mechanism itself in real neurons is clearly not responsible for the variabil- ity. We explore ideas about possible sources of spike-train variability in chapter 5.
Some neurons fire action potentials in clusters or bursts of spikes that can- not be described by a Poisson process with a fixed rate. Bursting can be included in a Poisson model by allowing the firing rate to fluctuate to de- scribe the high rate of firing during a burst. Sometimes the distribution of bursts themselves can be described by a Poisson process (such a doubly stochastic process is called a Cox process).
1.5 The Neural Code
The nature of the neural code is a topic of intense debate within the neuro- science community. Much of the discussion has focused on whether neu- rons use rate coding or temporal coding, often without a clear definition of what these terms mean. We feel that the central issue in neural coding is whether individual action potentials and individual neurons encode inde- pendently of each other, or whether correlations between different spikes and different neurons carry significant amounts of information. We there- fore contrast independent-spike and independent-neuron codes with cor- relation codes before addressing the issue of temporal coding.
In vitro
(slice):
regular spiking pattern for constant current injection in slices the dendritic trees are almost completely cut.In vivo:
irregular spiking pattern for constant current injection or external stimulus.dendritic trees are in tact and the neuron is exposed to fluctuating stimulations from other neurons
Montag, 15. April 13
The role of inhibitory neurons, balanced networks:
Model calculations:
integrate-and-fire-neuron exposed to external noise ddtU(t) = ⌧1
o
n
Uext + ¯U + ⌘(t) U(t)o
‘fluctuating force‘ due to coupling to populations of excitatory and inhibitory neurons⌘(t)
h⌘i = 0 ⌦
⌘2↵
= Tnoise
firing rates of background excitatory and inhibitory neuronsfe fi U¯ ⇡ Weefe Weifi Tnoise ⇡ Weefe + Weifi
weak noise
strong noise
Response characteristic:
hf(Uext)i
Uext weak noise
strong noise
Spiking activity strongly influenced by noise
increased sensitivity, tunable response characteristic
The role of inhibitory neurons, balanced networks:
Up and down state?
Oscillations?
Binding:
binding by synchronization of spindle oscillations
Conclusion?
New techniques reveal a broad spectrum of exciting and detailed answers.