So here's the thing: consider a detector, made of several vertical strings, each hosting 60 sensors distributed equidistantly, as shown in the following picture (dark dots are sensors):
Particles will then flow through and produce pulses on each sensor that you can use. Informations include time of passing by, coordinates (stringnumber+position on string or cartesian coordinates), total charge, etc.
The final goal is to reconstruct angles of incoming particles and their energies. Though we first just consider a simple classification problem of finding out whether the particle came in form the north or south hemisphere (i.e. from the top or bottom direction of the detector.
As input values we use every timestamp of each sensor, their exact position hasn't to be passed as it will be encoded by the position of the timestamp in the input list.
The problem we have is with sensor that didn't detect any particle. Is it clever to insert their timestamp as float("Inf")? Leaving their time as zero is another option, but it will then be ambiguous with the sensor that triggered as first (having timestamps normalised to [0,1]).
I'd be glad to hear any hint from people who have already worked with timestamps in neural networks. Also, if you have any good idea on developing NN's for our future tasks please share them! The planned approach are convolutional (deep) neural networks, but we still have to think about how to encode sensor-positions in that irregular hexagonal shape.
