Modeling working memory and decision making using generic
Download
Report
Transcript Modeling working memory and decision making using generic
Rules for Information
Maximization in Spiking Neurons
Using Intrinsic Plasticity
Prashant Joshi & Jochen Triesch
Email: { joshi,triesch }@fias.uni-frankfurt.de | Web: www.fias.uni-frankfurt.de/~{joshi,triesch}
Synopsis
Neurons in various sensory
modalities transform the stimuli into
series of action potentials
The mutual information between
input and output distributions
should be maximized
Biological Evidence: V1 neurons in
cat and macaque respond with an
approximately exponential
distribution of firing rates
Synopsis
Intrinsic plasticity is the persistent modification of a
neuron’s intrinsic electrical properties by neuronal or
synaptic activity
It has been hypothesized that intrinsic plasticity plays a
distinct role in firing rate homeostasis and leads to an
approximately exponential distribution of firing rate
This work derives two gradient based intrinsic plasticity
rules with both rules leading to information maximization
Rule 1: Direct maximization of MI
Rule 2: Minimize the Kullback-Leibler divergence
between OP distribution and a desired exponential
distribution
IP is achieved by adapting the gain function of a neuron to
its input distribution
Outline
Intrinsic plasticity in biology
Computational theory and learning rules
Neuron model
IP Rule 1
IP Rule 2
Simulation Results
Conclusion
Intrinsic Plasticity in Biology
Trace eyelid conditioning
task (Disterhoft et. al. )
Recordings from CA1
pyramidal cells showed a
transient (~1-3 days)
increase in excitability
Figure from: W. Zhang, D. J. Linden. The other side of engram:
Experience-driven changes in neuronal intrinsic excitability. Nat.
Rev. Neurosc., Vol 4, pp. 885-900
Neuron Model
Stochastically spiking neuron with
refractoriness (Toyozumi et. al. )
τm = 10 ms, τrefr = 10 ms, τabs = 3 ms
With refractoriness
Without Refractoriness
IP Rule 1: Direct maximization of mutual information
Key Idea: To maximize
(1)
Equivalent to maximizing
(2)
Where,
(3)
Substituting (3) into (2) we get the term for maximization as:
(4)
Learning rule consist of a set of update equations for various parameters φ of the gain function
(5)
Rule 1 Update Equations
u 0 u 0 u 0
u
u u
u 0
MI
u
e
g
r0
u u0 r0
u
1
e
u
u
MI
g
Note that similar analysis for the term r0 leads to an update rule which will cause the
value of r0 to increase without any constraint, hence it is not included in the set of
update rules
IP Rule 2: Minimizing the KL Divergence
Key Idea: Minimize the KLD between fy(y) and the optimal exponential distribution fexp(y)
KL Divergence is defined as:
Learning rule consist of a set of update equations for various parameters φ of the gain
function
Rule 2 Update Equations
IP
r 0 r 0 r 0
g
1
r 0
r0
u 0 u 0 u 0
u
u u
g
r 0
r
0
u 0 1 1 e 1
u
IP
g
u u 0 r 0
r
0
1 1 e 1
u 1
u
u
IP
Results: Performance of IP Rule 1 (MI Max)
Inputs: 100 Spike trains,
Gaussian Dist. (mean = 25
Hz. SD = 5 Hz)
ηMI = 10-3, T = 10 min.
u 0
MI
u
e
g
r0
u u 0 rg
u 1
e
u
u
MI
0
Results: Performance of IP Rule 2 (KLD Min)
Inputs: 100 Spike trains,
Gaussian Dist. (mean = 30 Hz.
SD = 5 Hz)
ηIP = 10-5, µ= 1.5 Hz, T =
16.67 min.
r 0
IP
g
1
r0
g
r 0
u 0
1 1 e r 0 1
u
IP
g
u u 0 r 0
r
1 1 e 1
u 1
u
u
IP
0
Results: Convergence of Rule 2
η
Evolution of trajectories from 3 different initial conditions
IP
= 10-3, µ= 1.5 Hz
Results: Phase Plots
ηIP = 10-3, µ= 1.5 Hz
Pair-wise phase-portraits indicating
the flow field, while keeping the
third parameter constant
Results: Behavior for various IP dist.
η
Gaussian IP Dist: Mean = 30 Hz, SD = 8Hz
Uniform IP Dist: Drawn from [0,60] Hz
Exponential IP Dist: Scale parameter, β = 30 Hz
IP
= 10-3, µ= 1.5 Hz
Results: Adaptation to sensory deprivation
First half: Gaussian
(mean = 30 Hz, SD = 5
Hz)
Second half: Gaussian
(mean = 5 Hz, SD = 1
Hz)
Conclusions
Two simple gradient based rules for IP are presented
First rule used direct maximization of MI
Second rule minimizes the KLD between fy(y) and the optimal exponential distribution
fexp(y)
Adapt the gain function of a model neuron according to sensory stimuli
Valid approach for neuron models which have continuous and differentiable gain functions
Works for several different input distributions
Leads to exponential output distribution, firing rate homeostasis, and adapts to sensory
deprivation
References
1.
R. Baddeley, L. F. Abbott, M. Booth, F. Sengpiel, and T. Freeman. Response of neurons in primary and inferior temporal
visual cortices to natural scenes. Proc. R. Soc. London, Ser. B, 264:1775–1783, 1998.
2.
M. Stemmler and C. Koch. How voltage-dependent conductances can adapt to maximize the information encoded by neuronal
firing rate. Nature Neuroscience, 2(6):521–527, 1999.
3.
J. Triesch. Synergies between intrinsic and synaptic plasticity mechanisms. Neural Computation, 19:885–909, 2007.
4.
H. Beck and Y. Yaari. Plasticity of intrinsic neuronal properties in CNS disorders. Nat. Rev. Neurosc., 9:357–369, 2008.
5.
T. Toyoizumi, J.-P. Pfister, K. Aihara, and W. Gerstner. Generalized Bienenstock-Cooper-Munro rule for spiking neurons that
maximizes information transmission. Proc. Natl. Acad. Sci., 102:5239–5244, 2005.
6.
S. Klampfl, R. Legenstein, and W. Maass. Spiking neurons can learn to solve the information bottleneck problems and to
extract independent components. Neural Computation, in press, 2007.
7.
A. J. Bell and T. J. Sejnowski. An information maximization approach to blind separation and blind convolution. Neural
Computation, 7:1129–1159, 1995.
8.
W. Zhang and D. J. Linden. The other side of the engram: experience driven changes in neuronal intrinsic excitability. Nat.
Rev. Neurosc., 4:885–900, 2003.