2015 26 03, Gerhard
This is a memo of the main aspects I retained from the meeting on March 25 with Anders and Pawel.
(1) Our colleagues usually do not look at the fine structure of firing. They are interested only in more global characteristics of activation (density of spikes).
(2) The memory patterns are coded as populations of cells, i.e. (sufficient) activity in this population means that the pattern has been activated. If the network is presented with a stimulus, the respective cells are activated. We see this in the synchAlpha_100s data set, where only the populations trained to recognize patterns are activated (the first 16 populations). Stimulation of other populations (the “upper 15”) do not result in a change of state - they remain in ground state (with the exception of a bit of activity in the inhibitory cells slighly moduating the gound state as a reaction to the stimulation).
(3) The potential data contains mainly inhibitory information. When looking at the voltage data of single cells, many of them don’t even spike once in the whole 100s. So what we see in the correlation of the potential data is the dynamics of inhibitions, which is of course connected to that of the activity. It seems that the inhibitory cells perform an integration of the spiking activity. It would be interesting to know if a time constant for this integration is known (I asked Pawel about this per email today).
(4) The sliding correlation computation detects mainly dramatic differences in the data and smears those very much. If we want to see finer structures, we need to use a shorter window and possibily use other means to smooth the data when exciting the visualisaiton system. This seems especially relevant when looking at the ground state, where inhibition is very much “evening out” spontaneous firing of all cells, i.e. there are not dramatic changes in the signal (as compared to the inhibition signals caused by activation and which we see as low frequency pulses of regative polarity).
(5) To be found out: what is the difference between a correlation and a normalized dot product?