i've been trying to investigate what neural network type would work the best for my type of project, basically multiple datasets of matrices of only numbers, no images

i've tried to read from all sources from matlab but so far i can't really have an answer on what would be best to be used to create a neural network in which i can input multiple files of separated datasets in which each one has its own matrix of data (that i flip and rot90 because they are essentially flipped and inverted, possibly from openneuro's page format or creator https://openneuro.org/datasets/ds004504/versions/1.0.7 )

btw i already tried to look into convolutional, but as far as i read it only works with images and requires a lot of workaround to make it work with matrices of text/numbers

(the openneuro is made of 19 channels)

and i also have another dataset that involves the whole thing in one string that i then reshape into its 16 channels http://brain.bio.msu.ru/eeg_schizophrenia.htm

(after reshaping them)

[Matt J]

To recommend a network architecture, it is not enough to know what the input is. We need to know what the target output is as well. Also, what do you see as the difference between a matrix of numbers and an image?

[Ri go]

as a side note i want to also point that the -.set- files contain multiple variables as in like containing variable name, group, age, etc, and has the variable data which is the eeg reading, i want to know if that would also influence when doing a neural network

and well the difference is as far as i know the fact that for a convolutional network it uses this command that its only for images, and i haven't found if it can be used with just matrices of numbers

[Purvaja]

Hey @Ri go,

As @Matt J said, it depends upon what you expect as your output. Like are you performing classification or regression? Also I don't see why you can't apply CNNs here, CNNs can also be applied to 1D or 2D numeric matrices. For EEG, people often use 1D CNNs across the time dimension (channels × time). This avoids the need to treat EEG as a “fake image.”

Since EEG is a time series, recurrent networks are often a better fit. MATLAB supports sequenceInputLayer, which works directly with channel × time sequences, so you wouldn’t need to force the data into an image format.

• 1D CNNs – Good for directly classifying time-series. MATLAB has a nice example here: https://www.mathworks.com/help/deeplearning/ug/sequence-classification-using-1-d-convolutions.html
• LSTMs / BiLSTMs – Great for capturing temporal dynamics in EEG. Check out https://www.mathworks.com/help/deeplearning/ug/long-short-term-memory-networks.html and even this ECG-LSTM demo (similar workflow): https://blogs.mathworks.com/deep-learning/2018/08/06/classify-ecg-signals-using-lstm-networks