Dyskretne i niedyskretne modele sieci neuronów
Winfried JustDepartment of Mathematics, Ohio University
Sungwoo Ahn, Xueying WangDavid Terman
Department of Mathematics, Ohio State University Mathematical Biosciences Institute
A neuronal system consists of 3 components:
1) Intrinsic properties of cells
2) Synaptic connections between cells
1) Network architecture
Each of these involve many parameters and multiple time scales.
Basic questions:
• Can we classify network behavior?
• Can we design a network that does something of interest?
Outline of the talk
• Network connectivity and discrete dynamics
• Definition of discrete dynamics
• Reduction of ODE dynamics to discrete dynamics
• A small network suggests discrete model
• ODE models of single-cell dynamics
• Relation to other models of discrete dynamics
Single Cell
v’ = f(v,w)w’ = g(v,w)
A cell may be:• excitable or• oscillatory
Variable v measures voltage across membrane. It changes on fast timescale.
Variable w is called “gating variable” and roughly measures the proportion of open ion channels.It changes on slow timescale.
____ v – nullcline- - - - w - nullcline
v1’ = f(v1,w1) – gsyns2(v1 – vsyn)
w1’ = g(v1,w1)
s1’ = (1-s1)H(v1-)-s1
v2’ = f(v2,w2) – gsyns1(v2-vsyn)
w2’ = g(v2,w2)
s2’ = (1-s2)H(v2-)-s2
v1’ = f(v1,w1) – gsyns2(v1 – vsyn)
w1’ = g(v1,w1)
s1’ = (1-s1)H(v1-)-s1
v2’ = f(v2,w2) – gsyns1(v2-vsyn)
w2’ = g(v2,w2)
s2’ = (1-s2)H(v2-)-s2
Two Mutually Coupled Cells
Sometimes consider indirect synapses:
xi’ = εx(1-xi)H(vi-) - εxxi
si’ = (1-si)H(xi- x) - si
Introduces a delayin response of synapse
• Fast or slow. Depends on and
• Excitatory or Inhibitory. Depends on vsyn
Synapses may be:
s – fraction of open synaptic channelsH – Heaviside functiongsyn – constant maximal conductance
Empirical observations
When the dynamics of this system is simulated on the computer,one observes rather sharply defined episodes of roughly equal lengthsduring which groups of cells fire (reside on the right branch of the v-nullcline)together, while other cells rest (reside on the left branch of the v-nullcline).
Membership in these groups may change from episode to episode; a phenomenon that is called dynamic clustering.
Experimental studies of actual neuronal networks, such as the olfactory bulbin insects or the thalamic cells involved in sleep rhythms appear to show similar patterns.
This suggests that one could attempt to reduce the ODE dynamics to asimpler discrete model and study the properties of the discrete modelinstead.
Transient: linear sequence of activation
Period: stable, cyclic sequence of activation
Reduction to discrete dynamics
(1,6)
(4,5)
(2,3,7)
(1,5,6)
(2,4,7)
(3,6)
(1,4,5)Assume: A cell doesnot fire in consecutiveepisodes
Some other solutions
(1,6)
(4,5)
(2,3,7)
(1,5,6)
(2,4,7)
(3,6)
(1,4,5)Different transient
Same attractor
(1,3,7)
(4,5,6)
Different transientDifferent attractor
NetworkArchitecture
(1,2,5)
(4,6,7)
(2,3,5)
(1,6,7)
(3,4,5)
(1,2,7)
(3,4,5,6)
What is the state transition graph of the dynamics?
How many attractors and transients are there?
2 7
6 5
4
1
3
Network architecture
145
36
247156
237
2367 367
245 24
1245
13567
2457
257
27
13455
123467
13456
7
12345612457
1356
135
56
153
12347
23467
14567
24567
124562456
2347
27
4513
136
1367
23457
1456
456
137
1237
16
2357
1
6
12367
234567
123457
3456
127
346 3467
345
167
235
467
125
34 256712567
1257
1235146
1467
14 46 1235723567
2345
67
2
25
1346
134
1346712345
134567
1267
267
352467
12467356
1247
1723456
1457
236
457
126
357
246
1357
1246
3457
1236
1256
1347
12356
26
5712346 47 256
34712356
23
234
1567
4567
12
123
4
567
14567
1234
34567
123567
124
3567
147
2356
157
2346
Remarks
1) We have assumed that refractory period = 1
If a cell fires then it must wait one episode beforeit can fire again.
2) We have assumed that threshold = 1
If a cell is ready to fire, then it will fire if it received input from at least one other active cell.
For now, we assume that:• refractory period of every cell = p• threshold for every cell = 1
Discrete DynamicsStart with a directed graph D = [ VD, AD ] and integer p.
2 7
6 5
4
1
3
Discrete DynamicsStart with a directed graph D = [ VD, AD ] and integer p.
A state s(k) at the discrete time k is a vector: s(k) = [s1(k), …., sn(k)] where si(k) {0, 1, … ,p} for each i. (n = # cells)
The state si(k) = 0 means neuron i fires at time k.
Dynamics on the discrete network:
• If si(k) < p, then si(k+1) = si(k)+1• If si(k) = p, and there exists a j with sj(k)=0 and <j,i> AD, then si(k+1) = 0.• If si(k) = p, and there is no j with sj(k)=0 and <j,i> AD, then si(k+1) = p.
Two Issues
1) When can we reduce the differential equations model to the discrete model?
2) What can we prove about the discrete model? In particular, how does the network connectivity influence the discrete dynamics?
Reducing the neuronal model to discrete dynamics
Given integers n (size of network) and p (refractory period), can we choose intrinsic and synaptic parameters so that for any network architecture, every orbit of the discrete model can be realized by a stable solution of the neuronal model?
Answer:
- for purely inhibitory networks.No
Yes - for excitatory-inhibitory networks.
Post-inhibitory rebound
We will consider networks of neurons in which the w-nullcline intersects the left branch of the v-nullcline(s).
If such a neuron receives excitatory input, the v-nullcline moves up, if it receives inhibitory input, the v-nullcline moves down.
If two such neurons are coupled by inhibitory synapses, the resulting dynamics is known under the name post-inhibitory rebound.
Purely Inhibitory Network
cell 1
cell 2
cell 3
cell 4
C(0)
C(0)
C(1)
C(1) g = 0
C(2)
1 2
3 4
(1,2)
(3,4)Note that the distance between cellswithin each cluster increases.
Excitatory-Inhibitory Networks
E-cells
I-cells
excitation
inhibition
Formally reduce E-I network purely inhibitory network
E-cells
I-cells
excitation
inhibition
E-cell fires I-cells fire E-cells firedue to rebound
We can then define a graph on the set of E-cellsand define discrete dynamics as before.
More precisely:
The vertex set of the digraph consists of the numbers of all E-neurons. An arc <i,j> is included in the digraph if and only if there is some I-neuron x that may receive excitatory input from i and may send inhibitory input to j.
Rigorously reducing E-I networks to discrete dynamics
Assume:
• All-to-all coupling among I-cells
• Inhibitory synapses are indirect (slow)
• Suitable functions f and g
• The ODE dynamics is assumed to be the dynamics on the slow timescale; all trajectories move along the v-nullclines; jumps are instantaneous
Discrete vs. ODE models
Consider any such network with any fixed architecture and fix p,the refractory period. We can then define both the continuousneuronal and discrete models. Let P(0) be any state of the discrete model. We then wish to show that there exists a solution of the neuronal system in which different subsets of cells take turns jumping up to the active phase. The active cells during each subsequent episode are precisely those determined by the discrete orbit of P(0), and this exact correspondence to the discrete dynamics remains valid throughout the trajectory of the initial state. We will saythat such a solution realizes the orbit predicted by the discretemodel. This solution will be stable in the sense that there is aneighborhood of the initial state such that every trajectory thatstarts in this neighborhood realizes the same discrete orbit.
Main Theorem (Terman, Ahn, Wang, Just; Physica D, 237(3) (2008))
Suppose a discrete model defined by a digraph is given. Then therere are intervals for the choice of the intrinsic parameters of the cells and the synaptic parameters in the ODE model so that:
1. Every orbit of the discrete model is realized by a stablesolution of the differential equations model.
2. Every solution of the differential equations model eventuallyrealizes a periodic orbit of the discrete model. That is, ifX(t) is any solution of the differential equations model, thenthere exists T > 0 such that the solution {X(t) : t > T } realizes a periodic orbit or a steady state of the discrete model.
Strategy
Suppose we are given an E-I network.
Let s(0) be any initial state of the discrete system.
We wish to choose initial positions of the E- and I- cells so that the E-I network produces firing patterns as predicted by the discrete system.
g = 0
E-cells
J2
J0
J1
J2
J0
J1
We construct disjoint intervals Jk, k = 0,…,p, so that:
Let s(0) = (s1, ….., sn). Consider E-cells, (vi,wi).
Assume: If si(0) = k, then wi(0) Jk.
Then: T* > 0 such that if si(1) = k, then wi(T*) Jk. The only E-cells that fire for t [0,T*] are those with si(0) = 0.
p = 2
Generalized Discrete Dynamics
Start with a directed graph D = [ VD, AD ] and vectors ofintegers p = [p1, … , pn] and th = [th1, … , thn].
A state s(k) at the discrete time k is a vector: s(k) = [s1(k), …., sn(k)] where si(k) {0, 1, … ,pi} for each i. (n = # cells)
The state si(k) = 0 means neuron i fires at time k.
Dynamics on the discrete network:
• If si(k) < p, then si(k+1) = si(k)+1• If si(k) = pi, and there exists at least thi nodes j with sj(k)=0 and <j,i> AD, then si(k+1) = 0.• If si(k) = pi, and there are fewer than thi nodes j with sj(k)=0 and <j,i> AD, then si(k+1) = pi.
Rigorous analysis of discrete model
Start with a directed graph D = [ VD, AD ]pi = refractory period of neuron i
thi = threshold of neuron i n = # of vertices.
How does the expected dynamics of the discrete model depend on the density of connections?
We will study this question by considering random initial states in random digraphs with a specified connectionprobability.
Some Definitions
Let L = {s(1), …., s(k)} be an attractor.
Act(L) = { i: si(t) = 0 for some t} (the active set of L)
L is fully active if Act(L) = [n] = {1, … , n}
L is a minimal attractor if Act(L) Ø and, for each i Act(L), si(0), …., si(k) cycle through 0, 1, …, pi.
Let: MA = {states that belong to a minimal attractor} FAMA = {states that belong to a fully active minimal attractor}
Consider random digraphs: (n) = probability <i,j> AD for given <i,j>.
Theorem: Assume that pi and thi are bounded independent of n.
(i) If (n) = (ln n / n), then as n , with probability one, |FAMA|/|states| 1.
(ii) If (n) = o(ln n / n), then as n , with probability one, |MA|/|states| 0.
A phase transition occurs when (n) ~ ln n / n.
Just, Ahn, Terman; Physica D 237(24) (2008)
Autonomous sets
Definition: Let s(0) = [s1(0), …., sn(0)] be a state.
We say A VD is autonomous for s(0) if for every i A,
si(t) is minimally cycling (that is, si(0), si(1), …, si(t)
cycles through {0, …., pi}) in the discrete system that is
obtained by restricting the nodes of the system to VD.
Example: Active sets of minimal attractors are autonomous. Note that the dynamics on an autonomous set does not depend on the states of the remainder of the nodes.
The following result suggests that there exists another phase transition ~ C/n.
Theorem: Assume that each pi < p and thi < th. Fix (0,1). Then C(p, th, ) such that if (n) > C/n, then with probability tending to one as n , a randomly chosen state will have an autonomous set of size at least n. In particular, most states have a large set of minimally cycling nodes.
Just, Ahn, Terman; Physica D 237(24) (2008)
Numerical explorations
Current work in progress
When the connection probability is ~ 1/n, another phase transition occurs for the case when all refractory periods and all firing thresholds are 1. Below this phase transition, with high probability the basin of attraction of the steady state [1, … , 1] becomes the whole state space. We are investigating what happens for connection probabilities slightly above this phase transition. Theoretical results predict longer transients near the phase transition, and this is what we are seeing in simulations. One question we are interested in is whether chaotic dynamics would be generic for connection probabilities in a critical range.
Other ongoing research
• Can we generalize our first theorem to architectures where the connections between the I-cells are somewhat random rather than all-to-all?
• How to incorporate learning and processing of inputs into this model?
• Can we obtain analogous results for networks based
on different single-cell ODE dynamics?
Hopfield Networks
• Networks are modeled as digraphs with weighted arcs; weights may be positive or negative
• Each neuron has a firing threshold thi
• At each step, neurons are in state zero or one• The successor state of a given state is determined by
summing the weights of all incoming arcs that originate at neurons that are in state one. If this weight exceeds thi, the neuron goes into state one (fires), otherwise it goes into state zero
Hopfield vs. Terman Networks
• Hopfield networks don’t model refractory periods• Terman networks don’t allow negatively weighted
arcs• For refractory period p = 1, both kinds are examples
of Boolean networks• Dynamics of random Hopfield networks tends to
become more chaotic as connectivity increases• Random Terman networks may allow chaotic
dynamics only for narrow range of connectivity
Why am I interested in this?
• My major interests are centered around models of gene regulatory networks.
• These can be modeled with ODE systems; but Boolean and other discrete models are also being studied in the literature, with the (generally) unrealistic assumption of synchronous updating.
• Question: Under which conditions can we prove a correspondence between discrete and continuous models of gene regulatory networks as in our first theorem?