Chapter 3 Neuron Models: From Biological to Computational
1 Why
Neurons are fundamental components of spiking neural networks. Information processing relies on different characteristics of neurons. Therefore, in this chapter, we will explore different neuron models that have different characteristics. It should be noted that there is never the best model, only the most suitable model. So, never evaluate a neuron model without its application scenario.
2 Single Compartmental Modelling
How to build a neuron model? How many details should a neuron model has? The compartmental modelling approach gives an answer to these questions.
Neurons consist of many components, including axons, dendrites, soma and etc. Some structures are very long in length and very large in number. It is very intricate to simulate a neuron as a whole. But we usually solve a complicated problem by simplifying it: split neurons into discrete areas and describing each discrete part separately is easier. So, such discrete parts are called compartments.
Then, we must face a new question: “How detailed should compartments be?”. Of course, if we have enough time and efforts, we can make compartments as small as possible. But this will lead to thousand of equations and parameters to construct a neuron model which greatly undermines efficiency of the model. Moreover, in spiking neural networks, we are more interested in neurons as a whole rather than spare no efforts to simulate neurons as detailed as possible. Taking these into consideration, we usually treat a neuron as one compartment so this is called single compartmental modelling. Figure shows different levels of compartments in neurons.
Fig. 1: Levels of compartmental models. (a) show the original neuron. (b) is detailed compartmental model. © is reduced compartmental model. Compared with (b), compartments in © represent larger areas. (d) is single compartmental model. The neuron is regarded as a whole.
Although detailed compartmental models can simulate neurons in an almost realistic way, they also lead to huge computational complexity. Single compartmental models simplify the representation of neurons. Such simplification makes these models easy to apply in large networks due to their balance between accuracy with complexity.
Therefore, in this chapter, we will introduce single compartmental models. They can be categorized into three groups. You can see them developing from biologically explainable to abstract, from biological to computational and from complex to simple.
3 Biologically Explainable Models
These models are based on ion channels on membrane. In biology, the ion flow between membrane triggers changes of membrane potential and then lead to spikes. The mechanism in biology is quite complicated so that here we only provide brief introductions to get you some rough understandings.
3.1 Hodgkin-Huxley Model
Hodgkin et al found by experiments that there were two main ion channels on membrane. One was sodium ion channel and the other was potassium ion channel. Apart from two main channels, they believed that there were channels mainly for chloride ion and formed the leakage current. They showed such mechanism by an equivalent circuit in Figure 2.
Fig. 2: An equivalent circuit of Hodgkin-Huxley model.
To describe its mechanism, we use the following differential equation system to specify how variables change with time.
{
C
m
d
V
d
t
=
−
g
L
(
V
−
E
L
)
−
g
ˉ
N
a
m
3
h
(
V
−
E
N
a
)
−
g
ˉ
K
n
4
(
V
−
E
K
)
+
I
d
m
d
t
=
α
m
(
V
)
(
1
−
m
)
−
β
m
(
V
)
m
d
h
d
t
=
α
h
(
V
)
(
1
−
h
)
−
β
h
(
V
)
h
d
n
d
t
=
α
n
(
V
)
(
1
−
n
)
−
β
n
(
V
)
n
\begin{cases} C_m\frac{dV}{dt}=-g_L(V-E_L)-\bar{g}_{Na}m^3h(V-E_{Na})-\bar{g}_Kn^4(V-E_K)+I\\ \frac{dm}{dt}=\alpha_m(V)(1-m)-\beta_m(V)m\\ \frac{dh}{dt}=\alpha_h(V)(1-h)-\beta_h(V)h\\ \frac{dn}{dt}=\alpha_n(V)(1-n)-\beta_n(V)n\\ \end{cases}
⎩⎪⎪⎪⎨⎪⎪⎪⎧CmdtdV=−gL(V−EL)−gˉNam3h(V−ENa)−gˉKn4(V−EK)+Idtdm=αm(V)(1−m)−βm(V)mdtdh=αh(V)(1−h)−βh(V)hdtdn=αn(V)(1−n)−βn(V)n
where
C
m
C_m
Cm is membrane capacity,
g
L
g_L
gL is leaky electrical conductance,
E
L
E_L
EL is leaky reversal potential (or resting potential),
E
N
a
E_{Na}
ENa and
E
K
E_K
EK are reversal potential of
N
a
+
{Na}^+
Na+ and
K
+
K^+
K+ respectively,
g
ˉ
N
a
\bar{g}_{Na}
gˉNa and
g
ˉ
K
\bar{g}_K
gˉK are max electrical conductance of
N
a
+
{Na}^+
Na+ and
K
+
K^+
K+ respectively. In above system, there are some variables that need to be defined further.
{
α
m
(
V
)
=
0.1
25
−
V
e
25
−
V
10
−
1
β
m
(
V
)
=
4
e
−
V
18
α
h
(
V
)
=
0.07
e
−
V
20
β
h
(
V
)
=
1
e
30
−
V
10
+
1
α
n
(
V
)
=
0.01
10
−
V
e
10
−
V
10
−
1
β
n
=
0.125
e
−
V
80
\begin{cases} \alpha_m(V)=0.1\frac{25-V}{e^{\frac{25-V}{10}}-1}\\ \beta_m(V)=4e^{-\frac{V}{18}}\\ \alpha_h(V)=0.07e^{-\frac{V}{20}}\\ \beta_h(V)=\frac{1}{e^{\frac{30-V}{10}}+1}\\ \alpha_n(V)=0.01\frac{10-V}{e^{\frac{10-V}{10}}-1}\\ \beta_n=0.125e^{-\frac{V}{80}}\\ \end{cases}
⎩⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎧αm(V)=0.1e1025−V−125−Vβm(V)=4e−18Vαh(V)=0.07e−20Vβh(V)=e1030−V+11αn(V)=0.01e1010−V−110−Vβn=0.125e−80V
As shown above, the Hodgkin-Huxley model needs very complicated definitions. Such complexity hinders its applications in spiking neural networks when efficiency is considered.
3.2 Morris-Lecar Model
Morris et al proposed a simpler model by simulating C a 2 + {Ca}^{2+} Ca2+ and K + K^+ K+ in neurons. The equivalent circuit of Morris-Lecar model are depicted by Figure 3.
Fig. 3: An equivalent circuit of Morris-Lecar model.
They applied two differential equations to describe mechanisms:
{ C m d V d t = − g L ( V − E L ) − g ˉ C a m ∞ ( V ) ( V − E C a ) − g ˉ K n ( V − E K ) + I d n d t = λ ( V ) ( n ∞ ( V ) − n ) \begin{cases} C_m\frac{dV}{dt}=-g_L(V-E_L)-\bar{g}_{Ca}m_\infty(V)(V-E_{Ca})-\bar{g}_Kn(V-E_K)+I\\ \frac{dn}{dt}=\lambda(V)(n_\infty(V)-n)\\ \end{cases} {CmdtdV=−gL(V−EL)−gˉCam∞(V)(V−ECa)−gˉKn(V−EK)+Idtdn=λ(V)(n∞(V)−n)
where I C a I_{Ca} ICa, I K I_K IK, I L I_L IL and I I I represent C a 2 + {Ca}^{2+} Ca2+ current, K + K^+ K+ current, leaky current and input current respectively and V V V represents membrane potential. Some parameters are defined further by following equations:
{ m ∞ ( V ) = 0.5 ( 1 + t a n h ( V − V 1 V 2 ) ) n ∞ ( V ) = 0.5 ( 1 + t a n h ( V − V 3 V 4 ) ) λ ( V ) = λ ˉ c o s h ( V − V 3 2 V 4 ) \begin{cases} m_\infty(V)=0.5(1+tanh(\frac{V-V_1}{V_2}))\\ n_\infty(V)=0.5(1+tanh(\frac{V-V_3}{V_4}))\\ \lambda(V)=\bar{\lambda}cosh(\frac{V-V_3}{2V_4}) \end{cases} ⎩⎪⎨⎪⎧m∞(V)=0.5(1+tanh(V2V−V1))n∞(V)=0.5(1+tanh(V4V−V3))λ(V)=λˉcosh(2V4V−V3)
4 Biologically Similar Models
Compared with above models, these models become simpler by giving up to describe hydronium between membrane. Instead, they directly analyse membrane potential. In this way, models can get rid of too many differential equations. They can be efficient enough to be applied in large networks but are still complicated due to non-linear differential equations.
4.1 FitzHugh-Nagumo Model
FitzHugh-Nagumo model simplifies Hodgkin-Huxley model by approximately treating n n n and h h h in Hodgkin-Huxley model as one constant. This simplification makes it possible to describe neuron in the following way:
{ d V d t = V ( V − α ) ( 1 − V ) − U + I d U d t = ϵ ( V − γ U ) \begin{cases} \frac{dV}{dt}=V(V-\alpha)(1-V)-U+I\\ \frac{dU}{dt}=\epsilon(V-\gamma U)\\ \end{cases} {dtdV=V(V−α)(1−V)−U+IdtdU=ϵ(V−γU)
where V V V is membrane potential, U U U represents restoring ability, I I I represents input current, α \alpha α, ϵ \epsilon ϵ and γ \gamma γ are all constants and 0 < α < 1 0<\alpha<1 0<α<1, ϵ ≪ 1 \epsilon \ll1 ϵ≪1.
4.2 Hindmarsh-Rose Model
Rose et al described neurons by the following system:
{
d
V
d
t
=
U
+
ϕ
(
V
)
−
W
+
I
d
U
d
t
=
ψ
(
V
)
−
U
d
W
d
t
=
ϵ
(
s
(
V
−
V
R
)
−
W
)
\begin{cases} \frac{dV}{dt}=U+\phi(V)-W+I\\ \frac{dU}{dt}=\psi(V)-U\\ \frac{dW}{dt}=\epsilon(s(V-V_R)-W)\\ \end{cases}
⎩⎪⎨⎪⎧dtdV=U+ϕ(V)−W+IdtdU=ψ(V)−UdtdW=ϵ(s(V−VR)−W)
where
I
I
I is input current,
ϵ
\epsilon
ϵ and
s
s
s are scaling parameters.
ϕ
(
V
)
\phi(V)
ϕ(V) and
ψ
(
V
)
\psi(V)
ψ(V) are functions that are defined by experience:
{
ϕ
(
V
)
=
−
a
V
3
+
b
V
2
ψ
(
V
)
=
c
−
d
V
2
\begin{cases} \phi(V)=-aV^3+bV^2\\ \psi(V)=c-dV^2\\ \end{cases}
{ϕ(V)=−aV3+bV2ψ(V)=c−dV2
4.3 Izhikevich Model
Izhikevich et al proposed a two-dimensional system to describe neurons:
{
d
V
d
t
=
0.04
V
2
+
5
V
+
140
−
U
+
I
d
U
d
t
=
a
(
b
V
−
U
)
\begin{cases} \frac{dV}{dt}=0.04V^2+5V+140-U+I\\ \frac{dU}{dt}=a(bV-U)\\ \end{cases}
{dtdV=0.04V2+5V+140−U+IdtdU=a(bV−U)
and if the membrane potential becomes no less than 30mV, the following reset mechanism will be applied:
{
V
←
c
U
←
U
+
d
\begin{cases} V\leftarrow c\\ U\leftarrow U+d \end{cases}
{V←cU←U+d
where
V
V
V is membrane potential,
U
U
U represents restoring ability,
a
a
a,
b
b
b,
c
c
c,
d
d
d are all hyper-parameters to control how neurons should fire.
5 Biologically Abstract Models
These models directly treat spikes as an event requiring some conditions. Non-linear differential equations are replaced with linear equations so that analysing is easier. For this reason, such models have been widely used in many experiments.
5.1 Integrate-and-Fire Model
Integrate-and-fire model directly describe the relationship between input current I I I and membrane potential V V V by the following equation:
τ m d V d t = V r e s t − V + R m I \tau_m\frac{dV}{dt}=V_{rest}-V+R_mI τmdtdV=Vrest−V+RmI
where τ m = C m R m \tau_m=C_mR_m τm=CmRm is time constant.
The reset mechanism is also simple: if membrane potential V V V goes equal or larger than threshold V t h r e s h o l d V_{threshold} Vthreshold, neurons will immediately generate a spike and the membrane potential V V V is instantly reset as resting potential V r e s e t V_{reset} Vreset which keeps unchanged during refractory period t r e f t_{ref} tref; otherwise the membrane potential V V V will gradually decrease to resting potential V r e s t V_{rest} Vrest.