Activation Function
- modify activation function in neurons.
Astrocyte Energy Regeneration
- accumulation rate //test
- Option to regenerate as wave function over time.
- if it’s too high, they would never run out
- if it’s too low, the signals would still die out
- modify astrocyte initialized energy (currently set at max).
- random between min and max provided values
Threshold
-
each neuron has to receive a certain # of signals to fire.
-
build-up of signals has to equal firing threshold.
-
non-negative value that stores # of signals received = acc.
- multiplies 0.9 with acc (decays at a constant rate)
- adds 1/6 if a signal is received (enough to outweigh decay)
- when it hits the firing threshold it will fire
- adds 1/6 if a signal is received (enough to outweigh decay)
- multiplies 0.9 with acc (decays at a constant rate)
-
refractory period = 5 seconds (default), can be modified.
m ← max(m + signals(now) - 0.1, 0) // decay cannot make value less than 0
if m >= 3 && enough time has passed // refractory period
then
m ← 0
signal
- should result in less signals
- calculate average signals received per neuron
- threshold & refractory period can be modified in real time.
Weights
-
Human Like Connectivity
- neuron randomly chooses their number of connections from a power-law distribution.
- few neurons have many connections, most neurons have few connections.
- what happens to max distance allowed for connection?
- graphical interface allowing to specify the distribution.
- make connectivity look like human brain connectivity matrices.
- primary sensory inputs, motor (decision) outputs.
- one directional connection.
- neuron randomly chooses their number of connections from a power-law distribution.
-
randomly assign weights to connections.
- modify above to
- mi ← max(mi + SUM0,connections.length(wi*signalsi(now)) - 0.1, 0)
- modify above to
-
initialization
- eventually: interaction of weight and distance between neurons
- eventually: bell curve probability of weights
-
0-5-10 probability increases - higher at 5 then 0 and 10
Inhibitors
- 20% of neurons are inhibitors
- inhibitors can have a larger weight value than excitatory
- 0-p of positive and 0-n of negative and possibly modify these values
- neuron has boolean: inhibitor true/false -eventually: colorful difference between inhibitory and excitatory signals
STDP
- implement Spike-timing dependent plasticity, and a learning task. -reinforcement-stdp?
- update connection weights, based on firing patterns.