[go: nahoru, domu]

Skip to content

This repository includes the code for "Input-modulation as an alternative to conventional learning strategies" by Esin Yavuz and Thomas Nowotny (presented at the ICANN 2016 conference, DOI:10.1007/978-3-319-44778-0_7). The code is based on the GeNN framework.

Notifications You must be signed in to change notification settings

esinyavuz/Input-Modulation-Learning

Repository files navigation

This repository includes the model used in "Input-modulation as an alternative to conventional learning strategies" by Esin Yavuz and Thomas Nowotny, accepted for the ICANN 2016 conference.

It consists of a model of plasticity in the honeybee olfactory system, modulated by appetitive conditioning. The model is simulated using the GeNN framework (Yavuz et al. 2016).

Neuron and non-plastic synapse models are based on Nowotny et al. 2013, where more information can be found. Mechanisms for input-modulation by reinforcement learning are explained in Yavuz and Nowotny 2016.

Prerequisites:

  • CUDA Toolkit (tested on v 6.5 and 7. The version of GeNN used in this paper won't work with v 8.0 because of the redefinition of the atomicAdd function for the double precision).
  • GeNN v2.1.1 (available here: https://github.com/genn-team/genn/tree/2.1.1)

#Neuron Parameters

##ORN Parameters

Parameter Value
t_spike (spike width) 0.1 ms
t_refract (refractory period + spike with) 0.2 ms
V_rest (resting potential) -60.0 mV
V_spike (potential at top of spike) 50.0 mV
b_rate (base firing rate) 3 Hz
l_max (maximum rate allowed) 50.0 Hz
\alpha (rate of adaptation) 4 Hz
\beta (rate of recovery from adaptation) 2 Hz

##PN Parameters

Parameter Value
g_Na (Na conductance) 7.15 \mu S
E_Na (Na equipotential) 50.0 mV
g_K (K^+ conductance) 1.43 \mu S
E_K (K equipotential) -95.0 mV
g_l (leak conductance) 26.72 nS
E_l (leak equipotential) -63.563 mV
C (membrane capacitance density) 143 pF
I_0 (bias current) 0.0 A

##LN Parameters

Parameter Value
g_Na (Na conductance) 7.15 \mu S
E_Na (Na equipotential) 50.0 mV
g_K (K^+ conductance) 1.43 \mu S
E_K (K equipotential) -95.0 mV
g_l (leak conductance) 26.72 nS
E_l (leak equipotential) -63.563 mV
C (membrane capacitance density) 143 pF
g_M (M conductance) 40 nS
kM_\alpha (rise rate for M activation) 25 Hz
kM_\beta (fall rate for M activation) 1.0 Hz
I_0 (bias current) 0.0 A

##LHI Parameters

Parameter Value
g_Na (Na conductance) 7.15 \mu S
E_Na (Na equipotential) 50.0 mV
g_K (K^+ conductance) 1.43 \mu S
E_K (K equipotential) -95.0 mV
g_l (leak conductance) 26.72 nS
E_l (leak equipotential) -63.563
C (membrane capacitance density) 143 pF
g_M (M conductance) 60 nS
kM_\alpha (rise rate for M activation) 80 Hz
kM_\beta (fall rate for M activation) 1.0 Hz
I_0 (bias current) -80.0 pA

#Synapse Paremeters

##ORN-PN synapses (non-plastic)

Parameter Value
g (conductance) 0.021 nS
\sigma_g (jitter for initialising g) 0.05
E_rev (reversal potential) 0.0 mV
\beta (postsynaptic decay rate) 20 Hz

##ORN-PN synapses (plastic)

Parameter Value
g_ini (initial value for conductance) 0.021 nS
E_rev (reversal potential) 0.0 mV
\beta (postsynaptic decay rate) 20 Hz
g_max (maximal conductance) 0.045 nS
\tau_{decay} (decay time for recovery) 300 s
\g_mid (medium value for conductance) 0.021 pS
\g_slope (slope of the STDP curve) 0.021 nS
p_0 (baseline for eligibility curve) -5e-12
\tau_p (time constant of eligibility decay) 1.5 s
A (cste to be added to p for STDP) 6e-14
\tau_{STDP} (time constant of STDP decay) 12 ms

##ORN-LN synapses (non-plastic)

Parameter Value
g (conductance) 0.012 nS
\sigma_g (jitter for initialising g) 0.1
E_rev (reversal potential) 0.0 mV
\beta (postsynaptic decay rate) 20 Hz

##PN-LN synapses (non-plastic)

Parameter Value
g (conductance) 0.3 nS
\sigma_g (jitter for initialising g) 0.1
E_rev (reversal potential) 0.0 mV
\beta (postsynaptic decay rate) 20 Hz

##LN-PN synapses (non-plastic)

Parameter Value
g (conductance) 1 nS
E_rev (reversal potential) -80.0 mV
\beta (postsynaptic decay rate) 10 Hz

##LN-LN synapses (non-plastic)

Parameter Value
g (conductance) 1.2 nS
E_rev (reversal potential) -80.0 mV
\beta (postsynaptic decay rate) 10 Hz

##PN-LHI synapses (non-plastic)

Parameter Value
g (conductance) 0.017 nS
\sigma_g (jitter for initialising g) 0.05
E_rev (reversal potential) 0.0 mV
\beta (postsynaptic decay rate) 10 Hz

#Reward parameters

Parameter Value
R_{0} (baseline reward -- or -1 x expectance ) -8.0
R_max (value of given reward) 40.0
\tau_{R} (Time constant of reward rise/decay) 50 ms

REFERENCES

About

This repository includes the code for "Input-modulation as an alternative to conventional learning strategies" by Esin Yavuz and Thomas Nowotny (presented at the ICANN 2016 conference, DOI:10.1007/978-3-319-44778-0_7). The code is based on the GeNN framework.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published