Brain Criticality - Optimizing Neural Computations
Summary
TLDRThe video script delves into the critical brain hypothesis, which likens the brain's neural networks to a system at the brink of phase transition, akin to Maxwell's equations for electromagnetism. It explains phase transitions, critical points, and their relevance to neuroscience, suggesting that the brain operates near a critical point to optimize information processing. The script explores concepts like neuronal avalanches, power laws, and the balance between excitation and inhibition, highlighting the brain's computational efficiency at the critical state.
Takeaways
- π§ Understanding the brain's complexity is a major challenge in neuroscience, with a goal to develop a comprehensive theory similar to Maxwell's equations for electromagnetism.
- π The critical brain hypothesis has gained attention; it suggests that neural networks operate near a phase transition point, akin to the critical state of water molecules coexisting in liquid and gas phases.
- π The concept of phase transitions is introduced with the example of water boiling, illustrating how energy can lead to a change in the macroscopic properties of a system.
- π Phase transitions are characterized by an order parameter and driven by a control parameter, with first-order transitions showing a discontinuity and second-order transitions being continuous.
- π‘ The Ising model, used to explain magnet properties, is highlighted as an example of a system undergoing a second-order phase transition, leading to a critical point with unique properties.
- π At the critical point, systems exhibit long-range communication and scale-free behavior, with self-similar patterns observable at all scales, a characteristic of fractals.
- π Neuronal avalanches, observed in neural cultures, are bursts of activity that spread across networks and are indicative of the brain operating near a critical point.
- π The distribution of features of neuronal avalanches follows a power law, suggesting a lack of characteristic scale in brain activity and supporting the idea of criticality in neural systems.
- π The balance between excitation and inhibition in the brain is crucial and can be adjusted to demonstrate phase transitions, with compounds that block these processes causing shifts from critical dynamics.
- π‘ Operating near a critical point is beneficial for the brain as it optimizes information processing and computational power, allowing for efficient transmission and reduction of uncertainty about inputs.
- π The study of criticality in neuroscience is an exciting and growing field, with potential applications in understanding brain function and developing treatments for neurological disorders.
Q & A
What is one of the Holy Grails of Neuroscience?
-One of the Holy Grails of Neuroscience is to build an elegant yet comprehensive theory that describes the general roles of the nervous system, similar to the impact of Maxwell's equations on electromagnetism.
What does the critical brain hypothesis propose about the operation of networks of neurons?
-The critical brain hypothesis proposes that networks of neurons operate near a point of phase transition, in a special critical state similar to when water molecules coexist in liquid and gaseous phases.
How does the phase transition concept relate to the boiling of water?
-The phase transition concept relates to the boiling of water as an example of a first-order transition, where water changes from a liquid to a gaseous state, absorbing latent heat and undergoing a qualitative change in properties.
What is an order parameter in the context of phase transitions?
-An order parameter is a macroscopic property that quantifies the organization of a system, such as entropy, volume, fluidity, or surface tension in the case of water.
How does the control parameter influence the phase transition?
-The control parameter, such as temperature or pressure, drives the phase transition by altering the system's state. It is an independent variable that can be freely varied in an experiment.
What is a continuous or second-order phase transition?
-A continuous or second-order phase transition is a type of phase transition where the change in the order parameter is continuous, allowing the system to smoothly transition from one phase to another.
What is the significance of the critical point in a phase transition?
-The critical point is a unique intermediate state at the interface of two different phases where the boundaries are blurred, and new properties emerge. It is a point of balance between order and disorder.
How does the Ising model relate to the concept of criticality in the brain?
-The Ising model, originally developed to explain the properties of magnets, demonstrates how a system can undergo a second-order phase transition, exhibiting properties of criticality such as long-range communication and scale-free behavior, which are relevant to understanding neural networks.
What is the role of the correlation length in the Ising model?
-The correlation length in the Ising model is the distance at which the dynamic correlation between spins drops to zero. It indicates the extent of long-range communication in the system, peaking at the critical point.
What is the significance of scale-free behavior in the context of the brain?
-Scale-free behavior signifies that the system, such as the brain, has no characteristic scale and resembles itself at any scale. This property is indicative of criticality and is associated with optimal information processing and computational power.
How does the branching ratio (Sigma) act as a control parameter in neural networks?
-The branching ratio (Sigma) is a control parameter that governs the transition from subcritical to supercritical dynamics in neural networks. When Sigma equals 1, the network is at a critical point where activity is neither decaying nor amplifying, and information transfer is optimized.
What experimental evidence supports the idea that the brain operates near a critical point?
-Experimental observations such as neuronal avalanches, which are power-law distributed patterns of activity in neural networks, support the idea that the brain operates near a critical point. These have been observed in various species and scales, from single neurons to EEG recordings.
Why is operating near a critical point considered beneficial for the brain?
-Operating near a critical point is considered beneficial for the brain because it maximizes capabilities such as information processing and computational power. At the critical point, the brain can efficiently transmit information and maintain a balance between activity and quiescence.
Outlines
π§ The Quest for a Neural Theory and Critical Brain Hypothesis
This paragraph introduces the complexity of understanding the brain and the pursuit of a comprehensive theory in neuroscience, akin to Maxwell's equations for electromagnetism. It presents the critical brain hypothesis, which suggests that neural networks operate near a phase transition point, drawing an analogy to the coexistence of water in liquid and gaseous states. The paragraph sets the stage for exploring the concept of criticality and phase transitions in the context of neural information processing.
π‘ Understanding Phase Transitions and Critical Points
The paragraph delves into the concept of phase transitions, using the example of water boiling to illustrate the qualitative change from liquid to gas. It explains order parameters and control parameters in phase transitions and introduces first and second order transitions. The critical point, where the system can exist in an intermediate state with unique properties, is highlighted as a point of interest for understanding the brain's operation.
𧲠The Ising Model and Its Relevance to Neuroscience
This section introduces the Ising model, originally developed for magnets, as a tool for understanding neural networks. It describes the model's lattice structure, where each site can be in one of two states, and how the alignment of these states can lead to macroscopic properties like magnetization. The paragraph explains how temperature acts as a control parameter, inducing fluctuations that can prevent the system from settling into the most energetically favorable configuration, and discusses the model's continuous second-order phase transition.
π Long-Range Communication and Correlation Length
The paragraph explores how long-range communication can emerge from local interactions in the Ising model, introducing the concept of correlation length to measure the coordinated behavior of distant sites. It discusses how dynamic correlation changes with temperature and how the system becomes scale-free at the critical point, with self-similar patterns observable at different scales.
π Power Laws and Scale Invariance in Critical Phenomena
This section discusses the power law distribution observed in the Ising model at the critical point, which indicates scale invariance. It uses a thought experiment to illustrate how the probability distribution of cluster sizes follows a power law, leading to the mathematical definition of scale invariance. The importance of power laws in understanding critical phenomena and phase transitions is emphasized.
π Neuronal Avalanches and the Brain's Critical State
The paragraph presents experimental evidence suggesting that the brain operates near a critical point, with neuronal avalanches observed in neural cultures. It describes how these avalanches resemble activity cascades in other systems and how they are characterized by power law distributions. The paragraph also discusses the potential implications of criticality for neural information processing across species and scales.
π The Branching Model and Neural Information Transmission
This section introduces the branching model to represent neural activity, where each neuron is connected to others with a transmission probability. It explains how the branching ratio, a control parameter, governs the system's behavior, leading to a phase transition at a critical value. The paragraph discusses how this model can be used to understand the balance between excitation and inhibition in the brain and the emergence of scale-invariant spike avalanches.
π― Optimizing Information Processing at the Critical Point
The paragraph explores why operating near a critical point is beneficial for the brain, maximizing its information processing capabilities. It uses a guessing game analogy to illustrate how the critical point optimizes information transfer, with the output layer of the branching model resembling the input layer's activity. The importance of this balance for neural computation and the potential applications of criticality research in neuroscience are highlighted.
π Conclusion and the Future of Criticality Research in Neuroscience
In conclusion, the video summarizes the key properties of systems undergoing second-order phase transitions, such as scale-free behavior and long-distance communication. It emphasizes the role of the critical point in optimizing the brain's information processing and the potential of criticality research for understanding brain function and developing treatments for neurological disorders. The video also promotes further exploration and discovery in this exciting field.
Mindmap
Keywords
π‘Neuroscience
π‘Critical Brain Hypothesis
π‘Phase Transition
π‘Order Parameter
π‘Control Parameter
π‘Second-Order Transition
π‘Icing Model
π‘Correlation Length
π‘Scale Invariance
π‘Power Law
π‘Neuronal Avalanches
π‘Branching Ratio
Highlights
Understanding the brain's complexity is a major challenge in neuroscience, with the goal of developing a comprehensive theory similar to Maxwell's equations for electromagnetism.
The critical brain hypothesis has gained attention, suggesting that neural networks operate near a phase transition point, akin to water's liquid and gaseous states coexistence.
The concept of phase transitions, such as water turning into gas, is used to explain the critical state in neural networks without actual melting or boiling.
Criticality in the brain is related to the efficient processing of neural information, with the phase transition being a significant point of interest.
Order parameters and control parameters are key in characterizing phase transitions, with the brain's critical point being driven by the balance of excitation and inhibition.
Second-order phase transitions, unlike first-order, involve continuous changes in the order parameter, allowing for the existence of a critical point with unique properties.
The Ising model, used to explain magnet properties, is also applied to understand critical points in neural networks, demonstrating long-range communication from local interactions.
Correlation length and dynamic correlation are measures used to understand how information is transmitted across neural networks at criticality.
Scale-free systems, like those at critical points, exhibit self-similarity at all scales and are characterized by power law distributions.
Neuronal avalanches, observed in neural cultures, are activity patterns that spread across networks and are indicative of critical brain dynamics.
Power law distributions of neuronal avalanches suggest the brain operates near a second-order phase transition, with no characteristic scale.
The branching model simplifies the complexity of neural networks, illustrating how activity can propagate and lead to critical phenomena.
The branching ratio, Sigma, is identified as a control parameter for neural networks, with a value of 1 indicating criticality and optimal information transfer.
Operating near a critical point maximizes the brain's information processing capabilities and computational power.
Criticality in neuroscience has potential applications in understanding brain function and developing treatments for neurological disorders.
The book 'The Cortex and the Critical Point' by John Max provides an in-depth look at criticality in neuroscience.
Brilliant.org is highlighted as an educational platform for learning computational methods and analysis of complex data sets like those in neuroscience.
Transcripts
understanding the brain in all of its
complexity is a difficult challenge one
of the Holy Grails of Neuroscience is to
build an elegant yet comprehensive
Theory describing the general roles of
the nervous system
similar to what a set of Maxwell
equations was to electromagnetism
in this video we are going to talk about
the critical brain hypothesis
a theory which has been getting a lot of
attention and experimental evidence in
recent years
it states that networks of neurons
operate near a point of phase transition
it is special critical State similar to
when water molecules coexist in liquid
and gaseous phases but what does it even
mean for the brain to undergo a phase
transition
there are surely no melting and boiling
right and in what way is this critical
point whatever it is important to neural
information processing all of that and
more coming right up
to understand the concept of criticality
let's begin with the notion of phase
Transitions and one such transition that
is probably very intuitive to you is
when liquid turns into gas consider a
pot of water on a hot stove at first
when the water is below 100 degrees the
energy provided by the burning stove
Heats it up accelerating individual
molecules at 100 degrees however
something interesting happens
even though the heat is still being
pumped into the water
temperature stays constant instead of
accelerating individual water molecules
the energy is now being spent on
Breaking the bonds between them
which allows the molecules to break free
from the lattice and fly away as a gas
notice that at the boiling point we have
seen a qualitative switch from liquid
which is incompressible has surface
tension and can dissolve substances to
gas which is compressible and doesn't
really dissolve anything even though the
individual water molecules stayed the
same the macroscopic properties of the
system as a whole changed drastically
what we have just witnessed is known as
the phase transition
more generally a phase transition is
observed when a system moves from
existing in one well-defined state of
organization or phase to another
phase is usually characterized by an
order parameter a macroscopic property
that somehow quantifies the organization
of the system for example in the case of
water order parameter can be entropy the
degree of disorder volume fluidity or
surface tension the phase transition is
driven by changes in control parameter
for example temperature or pressure
essentially you can think of the control
parameter as an independent variable
something that we can freely vary in an
experiment and the order parameter is a
dependent variable which changes in
response to altering the control
parameter and quantifies the state of
our system
notice that if you plot the order
parameter as a function of control
parameter then in the case of water
boiling you'll see a discontinuity a
sudden jump as we go from liquid to gas
accompanied by the absorbance of latent
energy such phase transitions are called
discontinuous or first order transitions
these are what you normally see in
everyday life however we will be more
interested in another type so called
continuous or second order transitions
as you might have guessed from the name
in this type of transitions the change
in the order parameter is continuous
which means that you can smoothly go
from one phase to another
now the slope of this curve can be very
steep sure but it's always continuous
this allows the system to exist in a
unique intermediate State called the
critical point right at the interface
when the boundaries between the two
different phases are blurred and new
properties emerge this is what makes the
critical point so special as we'll see
further
water can actually undergo a second
order transition at specific values of
temperature and pressure this is where
latent heat disappears and water can be
continuously transformed into gas
existing in a special state known as
supercritical fluid but to develop
intuition about the properties of
critical point let's focus on a
different much simpler system where the
second order transition feels more
natural and which actually has a lot of
applications in neuroscience
meet the icing model which was first
developed to explain the properties of
magnets the model consists of a large
lattice where each site can be in one of
two states plus one or -1 these sites
represent individual particles that the
system is made of characterized by their
spins which you can think about as this
sign of magnetic field generated by each
particle
when all spins are aligned tiny magnetic
fields of individual particles add up
resulting in the magnetic properties on
a macro scale
however when these pins are pointing in
different directions individual magnetic
fields essentially cancel each other out
and the system has no macro scale
magnetization
physicists have known for a long time
that if you heat up a magnet past a
certain point called Curie temperature
it will suddenly lose its magnetic
properties so there must be some sort of
phase transition at play now there are
two types of interactions that govern
the Dynamics of the system
first of all neighboring spins will tend
to line up because it is more
energetically favorable
we can write the energy of a single
pairwise interaction between the two
sides has been proportional to the
negative product of their spins where
the coefficient of proportionality J is
called the coupling constant and it
tells us how strongly the sides are
coupled as you can see when the spins
are of the same signs the energy of an
interaction is negative while for a pair
of opposite spins the value of energy is
positive
to find the energy of one site all we
need to do is to sum the energies of the
interactions between its four neighbors
and summing together the energies for
all these Heights will give us the
energy of the entire system
because it will try to minimize its
energy we can expect all these pins to
perfectly line up
so what can prevent this from happening
notice that so far we haven't included
our control parameter temperature in any
of the equations in reality however
flipping of spins is a stochastic
process which is subject to random
thermal movement
more formally the distribution of energy
values subject to random fluctuations
obeys the boltzmann distribution which
gets more broad at higher temperatures
but it's not that important to
understand phase transitions so don't
worry about it
essentially you can think of temperature
as this stochastic destabilizing force
that will induce fluctuations Jocelyn
spins around and preventing them from
settling down into their most
energetically favorable configuration
the higher the temperature the more
disordered the system would be
to quantify this let's introduce an
order parameter for example the
macroscopic magnetization which is
defined as the average value of spins
when the absolute value of magnetization
is big
it means that a large number of spins
are aligned and our system will display
magnetic properties let's play with the
temperature and see how macroscopic
properties of the system change at low
temperatures local interactions Dominate
and the system exists in this state when
a large portion of spins are lined up in
contrast if you heat it up the
temperature related fluctuations take
over all the spins are oriented randomly
and no macroscopic magnetic properties
are observed notice that as we turn the
temperature knob there is something that
looks like a very abrupt change in the
organization of spins from Euler to
disorder
indeed it turns out that icing model
undergoes a continuous second order
phase transition
in other words it is possible to catch
the system in an intermediate state
right at the point of critical
temperature where order and disorder are
perfectly matched and a number of
interesting properties emerge
before we move to the brain let's spend
a little more time exploring the icing
model
similarly to how neurons communicate
with each other across the entire brain
there is a crosstalk between spins in
our lattice
notice that if you flip once pin this
can make some of the neighboring spins
to flip which in turn affects pins two
sides away and so forth So eventually
even though only nearest neighbor
interactions are explicitly defined the
information about flipping one spin has
the potential to be transmitted across
the entire ladders
in other words long-range communication
emerges from purely local interactions
a good way to measure this is through a
quantity called correlation length for
example let's take two lattice sites
some distance apart and look at their
Dynamic correlation which is a measure
of how coordinated in time their
behavior is
more formally for sites I and J we can
express the dynamic correlation as
follows where angular brackets indicate
averaging over time the first term in
the parenthesis represents the amount by
which the spin I fluctuates from its
average value and likewise for
fluctuations of the spin J
notice that in order to maximize the
value of dynamic correlation the spins
should fluctuate in a coordinated
fashion for the product of the
parenthesis to remain positive let's see
what happens to the value of dynamic
correlation for a given pair as we
change the control parameter
at low temperatures the spins don't
fluctuate much so they are stuck at
their average values yielding a low
value of dynamic correlation
at high temperatures the spins fluctuate
wildly but they do so in a random
uncoordinated fashion at one point in
time the terms in the parenthesis might
be of the same sign while on another
occasion they might be of different
signs so on average the value of dynamic
correlation is low as well
at the critical temperature however
nearest neighbor interactions are
balanced by the thermal stochasticity
there is a coexistence of coordination
and fluctuation resulting in the large
value of dynamic correlation as spins
flip up and down in a coordinated way
so far we've only looked at the dynamic
correlation for a single pair of pins
and that of course will depend on the
distance between them let's now plot the
value of c as a function of distance
between the spins for different
temperatures
unsurprisingly it decreases with
distance because we've only Incorporated
local interactions into the model
and what's prominent is that the curve
corresponding to the critical
temperature decays much more slowly than
the two other ones for example we might
find that it extends up to 20 sites
before it falls to near zero let's turn
this distance at which the dynamic
correlation drops to zero as correlation
length
drawing correlation length as a function
of temperature reveals a sharp Peak at
the critical point right there at the
intersection of odor and disorder is
where a large distance communication
emerges if we were to make an analogy
this is the state in which neurons in
networks communicate most effectively
through the largest number of synapses
between them
another important property that emerges
at the critical point is that our system
becomes scale free to illustrate what
this means take a look at these three
magnified snapshots of the isig model at
three different temperatures exactly
critical slightly below and slightly
above the critical temperature can you
guess which is which well this is not so
obvious because they all look kind of
similar but let's zoom out to see the
whole lattice as we move further and
further away it becomes apparent that
the leftmost snapshot corresponds to a
sub-critical temperature because
everything is ordered
zooming out on the rightmost snapshot
reveals a great deal of disorder so this
must be this super critical State now
watch what happens as we zoom out when
the system is at the edge of the phase
transition no matter how close or how
far away we are things look very similar
in the ideal case right at the critical
point this pattern will continue to
Infinity in other words there is no
characteristic scale this
self-similarity when the system as a
whole resembles some of its parts at any
scale is a typical property of fractals
such zooming animations are surely
satisfying to look at but is there any
way we can objectively prove that this
is scale free while this is not
to mathematically describe the scaling
properties let's consider the
probability distribution of this size of
clusters that is connected domains of
the same spin
here are the examples of a few
relatively small clusters and a couple
of larger ones let's focus on the
critical case and build the intuition
for what this function f should look
like
suppose a dark wizard has shrunken you
to a random size so that you have no
idea how tall you are and trapped you in
the icing model at the critical
temperature the only way for you to get
out is to correctly guess the
probability distribution of cluster
sizes
so you grab a ruler and start walking
around measuring the size of every
cluster you see
let's say on average you find out that
every hundredths cluster you encounter
has the size of four square inches so
you can write down F of 4 equals 1 over
100 because the ruler has been shrunken
with you the exact units are irrelevant
what would matter in this case are
ratios so you begin looking for clusters
that are two times as large and discover
that the probability of observing an 8
square inches cluster is about 1 over
500 which is 5 times as large compared
to the probability of seeing a four
square inches one after repeating the
measurements for a few other cluster
sizes you discover an interesting
pattern every time you double the size
of a cluster its probability decreases
by a factor of five
now here is the crucial Point Let's
imagine that you were shrunken to a
completely different size for example
what if you were actually 10 times
taller what would you see in this case
well remember because the system is
right at the critical point it is scale
invariant
in other words no matter how small or
big you were you would see similar
pictures everywhere and hence similar
cluster distributions at any scale so
this ratio of the probability for
observing the cluster of size 2x versus
the cluster of size X would equal to
one-fifth at every scale
which means it doesn't depend on x
of course there is nothing special about
the factor of 2. we could have just as
well looked at any other ratio of
cluster sizes so let's substitute the
number 2 with a factor k
the value of this ratio will be also
different for different factors because
for instance the probability of
observing a cluster with a size 3x will
be different than the probability of
seeing a cluster of size 2x in general
we can write down the right hand side to
be some function J of K since as we have
already established it surely doesn't
depend on x
this right there is the mathematical
definition of scale invariance so
whatever the distribution of cluster
sizes is it should satisfy this
constraint it turns out that the only
function for which this equation holds
is of the form a times x to the power of
minus gamma where the number gamma is
called the exponent in that case J of K
is equal to K to the power of minus
gamma plugging in our values for cluster
sizes reveals that gamma equals 2.32
this behavior is called a power law and
it is central for the study of critical
phenomena you can easily see for
yourself that x to the power of minus
gamma is indeed scale free proving it
the other way around that any scale free
function should have this form is a bit
more tricky and requires taking a few
derivatives I won't go through the
derivation in this video but if you are
interested check out the references in
the description
another way to see why power laws are
scale free is to plot a power law
function for example x to the power of
minus 2.3 next to something similar but
which is not power law for example an
exponential function notice that if we
look at the graphs at different scales
the shape of the parallel curve stays
the same the only thing that changes are
the exact units if axis labels were
erased we couldn't distinguish them from
each other however the exponential curve
looks different at different scales and
hence is not scale free
usually such relations are plotted in
double logarithmic coordinates known as
log lock plots
where you take the logarithm of both
sides in log lock coordinates power laws
look like straight lines and the slope
of that line is related to the exponent
value gamma
importantly many other characteristics
of this system for example the value of
dynamic correlation we discussed earlier
all follow a power law distribution near
a critical point in fact when you see
that a lot of things are power law
distributed it often suggests that the
system is near a point of a second order
phase transition
now after we've developed an intuition
for many important Concepts in
criticality let's finally move to the
brain
[Music]
one of the first pieces of experimental
observations suggesting that the brain
might be operating near a critical point
came in 2003 when John Banks ended our
plans published their seminal paper
titled vironal avalanches in neocortical
circuits they grew cultures of neurons
taken from Red somatosensory cortex and
put them on an 8x8 grid of electrodes to
record spontaneous activity that arises
in this network and what they found was
a characteristic pattern of activity
spreading across the network in space
and time let's take a closer look at the
data collected by a single electrode it
records how extracellular voltage known
as local field potential changes with
time
if you have seen my video on theater
rhythms it is essentially the same thing
what is usually observed in these
cultures of neurons in a dish however is
that Baseline lfp occasionally shows
large deflections following a
characteristic shape of a negative Spike
superimposed on a positive envelope this
arises when a number of neurons in the
vicinity of electrode activate
simultaneously
we are going to treat such waveforms as
discrete events that are either present
or absent simply by finding when the
voltage Trace dips below a certain
threshold and also keep track of the
amplitude of every such event
thus our data will contain both a Time
point and an amplitude value for every
lfp event that caused the threshold
repeating this procedure for all the
electrodes gives us the pattern of
spontaneous electrical activity
spreading in the network
notice how periods of quiescence are
interspersed by what looks like chains
of activity arising reverberating in the
network and finally dying out
such bouts of activity were termed
neuronal Avalanches because they
resemble Cascades of activity
propagating in systems like pile of sand
or earthquakes
to Define an avalanche more formally
let's break up the entire time course
into short means and for each bin count
the number of electrodes that recorded
the lfp event for example it might go
like this
first there is no activity then one
electrode detects the Supra threshold
event
in the next time being another electrode
then two simultaneously then another one
then three others at the same time and
then activity dies out
this sequence of time bins with some
non-zero activity flanked by at least
one time being of quiescence at either
side is what we are going to call an
avalanche
and we can describe each Avalanche by
its duration as well as its size defined
as the number of electrodes that were
activated during the Cascade
note that in the alternative definition
size can also take into account the
amplitude values
surprisingly the authors found that all
these features of neuronal avalanches
were distributed according to a power
law revealing a prominent straight line
in log lock plots
this suggested that the network activity
had no characteristic scale and that the
brain might be operating near a point of
a second order phase transition
now all of this happened in 2003 and
since then Pablo described Avalanches of
activity were demonstrated in awake
behaving animals of many species
including worms zebrafish monkeys and
humans and on a variety of scales from
single neuron recordings to
electroencephalography
suggesting that criticality can indeed
be a universal phenomenon which
describes many aspects in neural systems
at this point in the video I think a
couple of pieces are still missing first
of all I've said that the critical point
is the intermediate stage of phase
transition but what are the distinct
phases for the case of the brain what
would be neural analogues to temperature
as the control parameter and
magnetization as the order parameter and
secondly why operating near a critical
point would be useful to the brain at
all
[Music]
to make the following discussion more
intuitive let's switch from thinking
about avalanches in terms of lfp events
on the electrode grid and view the data
as activity of individual neurons that
either fire an action potential or not
such switch is Justified because of
scale invariants remember we are going
to see similar Behavior at any scale
plus it has been established
experimentally that there are indeed
such Spike Avalanches whose features are
power law distributed
in this representation each circle is an
individual neuron rather than a single
electrode now it becomes easier to
interpret the activity propagation in
the network
whenever any run reaches its threshold
voltage it sends a signal to its
Downstream Partners increasing or
decreasing their probability of
subsequent firing
but because neurons connect to each
other in very intricate ways unlike the
icing model for example it is hard to
see who is connected to whom
let's make a slight simplification to
this network and rearrange the neurons
into a layered structure so that
information will always flow from left
to right
even though we don't allow any feedback
connections or Loops to a first
approximation it is a reasonable
assumption and importantly this
description known as branching model
still allows for avalanches and other
critical phenomena to emerge in the
branching model each neuron is randomly
connected to some neurons in the
downstream layer and each connection has
a transmission probability associated
with it
upon this Spike of a sender neuron the
connection will transmit a spike with a
certain probability and if this happens
on the next time point the receiver
neuron will activate and send the
information further
additionally each neuron has a very
small probability of spontaneously
activating even when it doesn't receive
any input this stochasticity provides
the network with some input drive to
work with
as you may have guessed how activities
spreads in the network will be mostly
controlled by the distribution of
transmission probabilities to tune this
with a single parameter let's introduce
a number Sigma called a branching ratio
and set the sum of our going
transmission probabilities for each
neuron to be equal to Sigma
for example if the branch in ratio is 1
this configuration of our going
connections is valid because the sum of
probabilities is equal to 1 and this is
not
by the way the branching ratio doesn't
really tell us the exact number of
connections for Sigma equal to 1 a
neuron can have a hundred connections
each with a probability of 100 or one
connection with the probability of one
the only thing that is constrained is
the overall sum of transmission
probabilities
it turns out that this branching ratio
is a control parameter for our system
and that a phase transition occurs when
Sigma is equal to 1.
to understand why this is the case
notice from the very definition of Sigma
it follows that it is actually equal to
the average number of descendants that
are activated per single active ancestor
neuron
let's run the simulation for different
values of branching ratio and see what
happens
as you can see when Sigma is equal to
0.5 any activity that rises quickly dies
out this is because on average it takes
two Upstream neurons to fire in order to
activate a single Downstream neuron
in a real brain that would correspond to
deep comatose state with very little
activity
increasing the value of Sigma 2 causes
the network to blow up with activity
similar to what you would see during
epilepsy this is because on average
single neuron activates two others which
makes the activity amplify with time
something interesting happens when Sigma
is equal to 1 however
in this configuration one neuron on
average activates one Downstream
descendant allowing the network activity
to be maintained on a relatively
constant level without decaying or
amplifying
now because this is a stochastic process
Avalanches will eventually die out but
their sizes and durations will be power
law distributed
this is why the branch in ratio is a
control parameter governing the
transition from decaying to amplifying
activity and we can use the average
density of active neurons as the order
parameter which is really low for
subcritical values of branches ratio but
quickly grows as soon as we hit the
critical point by the way in real
neurons this control parameter which is
equal to the average number of neurons
activated by a single Upstream neuron is
shaped by the balance between excitation
and inhibition which are the two
counteracting forces in the brain
Dynamics importantly because this
balance of excitation and inhibition is
something that we can control
experimentally it can further illustrate
that there is indeed a phase transition
at play namely adding compounds that
block inhibitory transmission
disrupted the power law distribution of
avalanches seen in the normal condition
leading to supercritical behavior
likewise adding compounds that block
excitatory synaptic transmission leads
to subcritical Dynamics again with the
disruption of power laws
okay by now the only left is why would
brain even evolve to operate near a
critical point is it actually useful it
turns out that at the critical point a
lot of brain's capabilities such as
information processing and computational
power are maximized to begin
understanding why this is the case let's
think of information transmission as a
guessing game
suppose someone provides an input to the
first layer of our branching model by
activating a random set of neurons and
our job is to correctly guess the number
of neurons that were activated by
observing only the output the activity
of the rightmost layer
for low values of branching ratio no
matter how many neurons were activated
in the beginning the activity quickly
vanishes before reaching the output
layer we are completely ignorant about
the input because there is no activity
Trace left
for high values of Sigma activity will
be Amplified causing the output layer to
be fully activated even for the weakest
input which makes the guessing again
very difficult
at the critical case when Sigma is equal
to 1 the connection strengths are tuned
to the optimal intermediate value
because one neuron on average activates
only one descendant activity of the
output layer has a high probability of
resembling the activity of the input
layer in terms of the number of active
units
in other words observing the output
successfully reduces our uncertainty
about the input
if we introduce some measure of
information transmission that will tell
us how accurately we can make this gas
this quantity will have a sharp Peak
when the branch in ratio is 1.
similar to how Dynamic correlation Peaks
at the critical temperature in the icing
model this on a very simplified level is
what it means to optimize information
transfer
of course we have really just scratched
the surface of the topic of critical
point and its relevance to Neuroscience
I didn't really talk about the concept
of universality and the Beautiful
relationship between the exponents the
idea of course a criticality and how the
brain actually maintains this right
balance in the first place but if you
are interested to know more about
criticality and face transitions in the
Neuroscience context I really encourage
you to check out a wonderful book
published in 2022 titled the cortex and
the critical point by John Max who is
one of the pioneers of this field the
book is written in a really accessible
language and introduces all the concepts
from the ground up
before we move to the summary I have one
important message
in this video we have seen a few
examples of how analyzing neural data
can help us better understand the brain
but have you ever wondered how computers
can be programmed to process and analyze
such complex data sets if you're
interested in the fascinating world of
computer science you are definitely
going to love our today's sponsor
brilliant.org brilliant is an Innovative
educational platform that takes a
Hands-On approach to learning whether
you are a student or a professional it
is the best way to learn and Excel in
stem Fields Brilliance courses are
designed to be engaged in an interactive
packed with stunning demonstrations and
exercises to help you gain intuition and
develop problem-solving skills in the
context of real world applications
brilliant has thousands of lessons for
everyone from foundational math and
logic to Ai and quantum mechanics with
new ones being added monthly if you are
interested in getting started with
computational Neuroscience take a look
at the course on algorithm fundamentals
introducing essential array algorithms
such as sorting and binary search that
underlie the majority of modern
Computing and from there you can move on
to programming with python which is what
I've actually used to run simulations
you've seen in this video don't hesitate
to start your Learning Journey today go
to brilliant.org artem carsonov to get a
30-day free trial of everything
brilliant has to offer additionally the
first 200 people to use this link will
get 20 off the premium subscription
alright let's recap in this video we
have explore
X properties that arise when the system
undergoes a second order phase
transition namely it becomes scale free
and there is an emergence of long
distance communication between the
components
in the brain such phase transition is
governed by the fine balance between
excitation and inhibition and by
hovering near a critical point our
brains optimize information processing
of course this area is still very new so
there is a lot of things we don't know
but it is now apparent that the
applications of criticality research are
far-reaching and exciting from
understanding the general principles of
brain function to developing new
treatments and therapies for
neurological disorders as our
understanding of criticality in neural
networks continues to grow it is sure to
be an exciting field of exploration and
discovering in the years to come
if you like the video share it with your
friends and colleagues subscribe to the
channel if you haven't already press
like button and consider supporting me
on patreon stay tuned for more
interesting topics coming up
goodbye and thank you for the interest
in the brain foreign
[Music]
5.0 / 5 (0 votes)