on Monday, October 16th, 2017 2:26 | by Christian Rohrsen
Video of the attractor projected from a chunk of flight trace. Here one can see the difference in the trajectories of up- and down spikes. So here is another way of spike sorting by the way :).
Coloured traces depending where the locate in the attractor projection. Green means that they lie outside and when they stay with the PCs around zero its in red. The fly1r above does not look so clean as the v8 fly below (the cleanest measure I have)
on Monday, October 9th, 2017 2:34 | by Christian Rohrsen
These are the results of the SMAP for the TNTxWTB. I also have done a few for the c105;;c232xWTB but there is not much to say. I would say that the cleanest lines show a bigger slope, but prone to subjectiveness.
on Monday, October 2nd, 2017 5:35 | by Christian Rohrsen
This is what I showed about one fly from this line showing the attractor.
This graph is what I forgot to show in the lab meeting. There are the 6 best traces from this same line. All of them selected ad hoc subjectively. The three best of them to my eyes are exactly the three above in the graph (v8-this one is the one shown in the picture with the attractors-,v4,v2). What does this means? the ones with better traces (subjectively) have higher offset in the phi, this means that they are more predictable overall (maybe because more resolution?). In addition, they show higher slopes which means more nonlinearity.
on Monday, April 4th, 2016 3:04 | by Christian Rohrsen
In the last meeting Björn proposed to do correlations of cumulative increasing bins. He said to do that taking the zeroth point (last library point where prediction is still not done) and use it for having a potential 1 of correlation coefficient at the beginning. I could not do that because I didnt save the zeroth points, and this will be a bit tedious and confusing considering that many flies were tested and probably the order is not 100% known. Thus, I just did the bins skipping this zeroth point. After all, we should see something similar with this one. First two graphs: c105;;c232>TNT (first and second prediction point), second: WTBxTNT, third: WTBxc105;;c232.
Examples of how each of the flies look like. So they are basically cumulative bins with each single fly (each in different colour). Just to have a hint how does the singularity looks like. Second thing I did is normalize the to have a range from -1 to 1 all of them (I have to double check the range in the script) and also setting them at a starting point of zero. I did this because we do not want to have differences in the correlation coefficient due to a different offset of the values of the wing beat and neither because of the starting point (if the fly was already flying to the right full gas, then it could be that it has an influence in the following prediction).
c105;;c232 –> first at starting at zero without normalizing and then with normalizing. The next is just the RMSE (not so important).
on Monday, March 21st, 2016 1:13 | by Christian Rohrsen
To see if there is an exponential decay in the prediction of the fly traces we did correlations of bins of 40 data points. We have 4 graphs (the last one merged) which consist of predictions at two different points with two different number of neighbours used for the prediction. So we have for each group sucesively: prediction at the first prediction point with the first number of neigbours, then the same with different number of neighbours. The last two are two different numbers of neighbours for the second prediction point. In the order: c105;;c232>TNT, TNTxWTB, c105;;c232xWTB
on Monday, March 14th, 2016 1:48 | by Christian Rohrsen
This is now the results from trying to predict the fly behavior doing ensembles of two predictions for the next 200 data points at two different points of the traces.
From what we see here, there is no “flattening” in the prediction of the fly when the neurons under c105 and c232 are targeted by TNT. This is done with around 14/15 flies for each group with two predictions in each ensemble of the two starting points. That makes a total of 15flies x 3 groups x 2 starting points for prediction x 2 predictions per ensemble = 180 prediction traces. Now I´m trying to calculate it by making correlations of bins in the prediction-observed for the same fly
on Monday, February 15th, 2016 2:25 | by Christian Rohrsen
This is the prediction analysis of flies in the platform under a 20min experiment under dark conditions. The number of experiments change drastically among groups because of technical problems: WTBxTNT is 4, WTBxc105;;c232 is 22, for the experimental line is 6 (c105;;c232>TNT), for the platform without flies is 10. I show the root mean squares and the correlation coefficient for each group.
This is the experimental group: c105;;c232>TNT.
The group without flies on the platform. I expect here to get a very good predictability overall:
on Monday, February 8th, 2016 12:11 | by Christian Rohrsen
These were done the week before last one but I could not upload it last time, so here are they. These are the results for predicting 4-5 flies of each group. Just one prediction from the middle of the time series for the 40 data points ahead in the future. First group is c105;;c232>TNT.
This is c105;;c232 x WTB
As I saw that the graph had so much zig-zag I told Pablo to make a bigger number of tested flies and this is what he is presenting today.
In addition I did analyze other parameters which are all saved under a PDF file below (Strokelitude). This contains some parameters with doubtfull processing which I still don´t trust so I have to find a better way for the calculation of it.
on | by Pablo Martinez
After we got the data of the behavior of the fly with strokelitude, we have made the analysis of 40 datapoints. Those datapoints were randomly picked in the middle of our dataset (10 minutes) and we made a second prediction, taking the points just before the end of the flight. For that we made a double correlation with those two predictions for each line and those are the results:
For the two control lines,
WTB x C105;;C232
UAS-TNT-E x WTB
And for the experimental line (UAS-TNT-E x C105;;C232)
As long as the linear regresion decreases, more impredictable is the fly so its behave like non-linear function. The two controls are suposed to be more impredictable than the control line. There is more decrease in those two lines comparing to the experimental line, although is not too remarkable.
We also made an RMSE analysis, Christian explained it in his post. “RMSE measures just the differences of the absolute points whereas correlation coefficient is rather if the direction and degree of variation correlates (covariates)”
Here we have the plots:
WTB x C105;;C232(control)
UAS-TNT-E x WTB(control)
experimental line (UAS-TNT-E x C105;;C232)
As long as the linear regresion increases, the absolute points differ from the prediction and the normal trace.
on Sunday, January 17th, 2016 7:02 | by Christian Rohrsen
These are just 5 flies (WTBxTNT) from the strokelitude where I measured the correlation coefficient on the Y-axis. In the X-axis, first bin is from 0-2 s of prediction, second is 2-4s and so on.
It seems as if some flies do nicer than others. Although it seems to me that a correlation coefficient from 0.3 isnt a big thing with all this variability. I have to find out the best binning though, I think it needs to be much more in the short term.
When I do the mean of the 5 flies measured, I do see a very slight decay. But once more I would say the decay is from the bin 1 to the second.
Here I tried another way, the RMSE, which according to literature and to my own reasoning should be a better analysis. I think RMSE measures just the differences of the absolute points whereas correlation coefficient is rather if the direction and degree of variation correlates (covariates). I find a very weird result. The fit is bad, the it gets better (but it should be just a chance event because correl coef decreases) and then it get very bad and so on.
I think for the future I have to make ensembles of two k neighbours maybe, which seem to increase the prediction power 10-15%. And maybe not look that much into the future as it was done here (10s).
Here some examples of predictions vs observations: