Non-image neural network

Hello,

I'm working on a concept for a science instrument where the detectors are read in strips to generate 1D projections (see below). We only get a set of projections for each side, either 0 degrees or 90 degrees. There are 4 projections (with pulse heights/amplitudes for each of the strips) with size are 35x1, so the input to the neural network would be something like 35x4 (or something with the info reshaped to whatever works).

I have simulated events for neutrons (see below for a general idea), where a single or double tracks are read. I need to train a neural network to do a few things

1) Classify event as single or double track.

2) If possible obtain location (start/end coordinates in X, Y and Z) for each of the tracks.

Any guidance on how to do this in Matlab will be greatly appreciated.

Georgie

Answer 1

[Krishna]

Hello George,

From my understanding you are trying to predict whether an event is single track or double track using sensor data. The format of sensor data is 35 x 4 as mentioned in the description, and you would like to predict the same using neural networks.

The choice of architecture depends on the question and the type of data. If the sensor data is time dependent or periodic in nature you could use “narxnet”, “narnet” or “timedelaynet”, for more complex time dependencies use RNN like LSTM, otherwise simple feedforwardnet would work. Documentations below might be a great starting point,

https://in.mathworks.com/help/deeplearning/ref/feedforwardnet.html

https://in.mathworks.com/help/deeplearning/ref/narxnet.html

https://www.mathworks.com/help/deeplearning/ref/narnet.html

https://www.mathworks.com/help/deeplearning/ref/timedelaynet.html

The input of the dataset can be of the format 140 x 1 (35 x 4 spread out along columns). Further as the input data and output data both are categorical in nature you can use label or one hot encoding depending on the need. (For ex: 0 degree is 0 and 90 degree is 1).

Obtaining locations for each of the tracks depend on if you have the required dataset or not. If you have the dataset, you can simply use a regressive neural network and predict the locations.

Hope this helps.