2015 02 18, Gerhard


This is a description of the data set synchAlpha_100s_spikes1.out, which corresponds to the set synchAlpha_100s_volts1.out described at this page. The latter contains a subset of the cells of the former and consists in recordings of the cell membrane potentials with a sampling rate of 1000 Hz.


The spikes data set only contains the times at which each cell in the network has fired. There are also cells which never fired in the 100 seconds of the simulation - they were probably not connected in the training phase of the network.


The spikes data set contains the spike times of 32768 cells in a time interval of 100 s. The precision of the time tags is 0.1 ms (10 kHz). The 32768 cells represent two networks of 16384 cells each. The two networks are refered to as C1 and C2, where C2 is connected to C1 by a forward-only connection (C1 => C2).

Of each of the 16384 cells, 15360 form the "actual" networks, which are structured in 16 hypercolums containing each 32 minicolumns of 30 cells (16 * 32 * 30 = 15360). The other 1024 (16384 - 15360) cells are special and not discussed here.

In the slideshow to the right, all cells for both networks are plotted, once in the sequence according to their indices (y-axis in the plot), i.e. ordered by hypercolumns and once ordered by minicolums. In the second set the excitations and their responses of the network can be seen clearly as dark areas. In the 100 seconds 50 excitations did occur.

The task the network is trying to accomplish is to recognize the 16 different patterns is has been trained for. The network is exposed to a pattern by exciting the respective minicolumn in each hypercolum, i.e. pattern 1 excites all minicolums 1 in all hypercolumns, pattern 2 all minicolumns 2, etc.. It can be seen from the plots, that the first 16 excitations (of the first 16 minicolumns in each hypercolum) produce a different response than the excitations 17-31 (only 15, sic!), to which the network doesn't seem to respond at all. If this signifies that the network has been trained only for 16 patterns and therefore other excitations (from 17 onwards) don't produce a significant response has yet to be clarified. Also it would be interesting to know why these excitations where performed at all.

From excitation 32 onwards (which excites again all first minicolumns) a second sequence of 16 pattern responses can be observed (+ 3 excitations with almost no response again at the end). I assume that these should be in some way comparable to the first sequence, so it could be an interesting goal to characterise the differences in the responses auditorily as a basis for comparing the first set of responses with the second set. If the network is assumed to perform equally well in both cases, i.e. detects the pattern ("remembers" it), then some kind of similarity should be audible.

Taking a closer look at the excitations reveals that only the minicolums of the first 8 hypercolumns are actually excited (c.f. slideshow to the right, which shows excitation 4 and 500 ms of the response in 3 different views).

What can also be seen in the case of all patterns is that after the excitation, the excited cells (i.e. cells of the minicolumns of the lower 8 hypercolumns) do not fire for some time. The cells in upper 8 minicolums seem to respond first and also continue their activity once the cells from the lower 8 hypercolumns respond. There doesn't seem to be any significant response from other cells in the network. After a second latest the response seems to be over. Some are significantly shorter.

The excitations have some random structure but also show a clear pitch when listened to, which is introduced by the smallest possible entry delay between spikes of 0.1 ms. In the audification (c.f. player to the right) all 50 excitations sound quite alike. As the excitation is only some 20 ms long, the audification has been slowed down by a factor of 10 (200 ms per excitation, 100 ms pause between excitations), which transposes the pitched component down from 10 kHz to 1 kHz.

The excitation start times have been determined "by hand" and can be found in the table to the right (use scroller to see all entries).

2015 02 24, Gerhard


It seems that mainly the cells in the excited minicolumns respond to the excitation. This suggests that corresponding minicolumns are conneted accross hypercolumns. Also, the non-excited minicolums produce less spikes than usual during the response of the excited minicolumns. This can be seen from the plot on the right side which shows spike counts for the 50 events of the excited minicolumns (top curve), of the other minicolumns (middle), and the total count (bottom). The two plots in the slideshow show the count for the first 500ms of the event and the first 1800 ms (which is roughly the durattion of the shortest response). The 1800 ms plot suggests (little difference in the total number of spikes with first 16 and next 15 excitations) that there is a constant spike density in the network and spikes are distributed differently if there is an excitation and response (event). This can also be seen in the excitations, where the counts in the excited and non-excited cells are complementary (if there are more spikes in the excited cells, the others fire less).

2015 02 25, Gerhard


The 4 soundfiles to the right contain sonifications of the spikes in the excited minicolumns for all 50 events. The all spikes in the 30 cells of each of the 16 minicolumn have been added up. The 16 minicolumns (being part of another hyperoclumn each) have been spatialised such that the lower 8 hypercolums come from 8 speakers left to the listener (arranged from back to front, along the side of the virtual hall) and the upper 8 from the right. Remember that the excitation happens only in the lower 8 hypercolumns, so will be heard from the left speakers).


The spikes in each minicolumns are slighly coloured by convolution kernels of different frequency (differentiated gauss curve). Each side corresponds to a diatonic scale, the right side being shifted up by a semitone. But the kernels are so short, that the pitches are no perceived as such but just as slightly different colors. As it is not clear to me yet, how different spikes should stream to a structure, I prefer this neutral version. I have experimented a lot with other solutions, leading nowhere so far. The first soundfile contains 1.8 s of each excitation and response, the second one only 0.5 seconds right after the excitation (so just the main part of the response) and the third and the fourth are the same than the second but streched by a factor of 8. The third contains the first 16 events and the fourth the second 16.


The 5th soundfile contains all not excited minicolums per event, i.e. the excited ones are taken away for each event (which is a different set of minicolumns for each event). Per event the first 1800 ms are taken and strung together without a pause in order to hear better when the density drops, which is when there is an excitation and response in the other cells. To keep in mind when comparing it to he responses in the ecited cells: the overall density in this case of about 31 times larger, because it is all other minicolumns combined.


The sixth soundfile contains a non-spatialised version of the fifth sped up by a factor of 100 and band-limited to about 60 Hz in the original signal (i.e. 6 kHz in the sound). It is basically noise with almost no structure on this time scale. Only a 50 Hz hum can be heard, which corresponds to the 0.5 Hz of the excitation frequency.


All files best be listened to with headphones, as binaural auralisation of speakers in the Ligeti hall in MUMUTH is applied (using StiffNeck).


Finally got around reading the paper, which helps a lot (but also raises new questions, a list of which I have compiled for my meeting with Pawel on Friday): Mikael Lundqvist,Pawel Herman, and Anders Lansner, "Effect of Prestimulus Alpha Power, Phase, and Synchronization on Stimulus Detection Rates in a Biophysical Attractor Network Model", The Journal of Neuroscience, July 17, 2013, 33(29):11817–11824

estimated start times of excitations in seconds

 1  0.665

 2  2.680

 3  4.525

 4  6.685

 5  8.625

 6 10.520

 7 12.555

 8 14.610

 9 16.690

10 18.695

11 20.535

12 22.695

13 24.690

14 26.595

15 28.660

16 30.530

17 32.585

18 34.685

19 36.660

20 38.690

21 40.630

22 42.505

23 44.670

24 46.690

25 48.635

26 50.650

27 52.650

28 54.575

29 56.630

30 58.535

31 60.640

32 62.510

33 64.555

34 66.510

35 68.520

36 70.665

37 72.640

38 74.565

39 76.690

40 78.505

41 80.585

42 82.575

43 84.655

44 86.660

45 88.535

46 90.595

47 92.590

48 94.630

49 96.640

50 98.650

50 events, first 1800 ms of the excited minicolumns, including excitation

50 events, first 500 ms of the excited minicolumns, excluding excitation

first 16 events, first 500 ms of the excited minicolumns, excluding excitation, stretched 8 times

second set of 16 events, first 500 ms of the excited minicolumns, excluding excitation, stretched 8 times

50 events, first 1800 ms of the all not excited minicolumns, no pauses between events

all not excited minicolumns, sped up by a factor of 100 and band limited to 60 Hz in the original