2
$\begingroup$

Spike sorting consists in clustering groups of signals based on their shapes.

I have always seen convolutional neural networks (CNNs) being used for image classification. Nevertheless, in this case, the inputs that need to be classified would be one-dimensional. Would this be a problem for a CNN to achieve good results?

There are some things that I'm not sure about this being a good approach.

  • The two-dimensional kernels would become one-dimensional FIR filters. For these to extract relevant features, shouldn't all the spikes in the signals be in the same position? Would the recognition be position (or time, as we are talking about signals) invariant?
  • What about the pooling layers? Do they make sense if we are talking about 1D patterns?
  • Would it be a better approach to make the inputs of the CNN be pictures of the signals? For instance, an image from an oscilloscope instead of an array containing the values the signal takes for each instant of time. I don't think this would be very robust, but it was just an idea that crossed my mind.
$\endgroup$
1
$\begingroup$

There are several works where 1D Convolutions are applied in a similar way.

E.g. Effective Use of Word Order for Text Categorization with Convolutional Neural Networks R. Johnson, T. Zhang https://arxiv.org/pdf/1412.1058.pdf

If this approach could make sense or sense, depends on your data. You should use a proper filter size for your convolutions. The same is for the pooling, you have to find the setup which best fit you data.

For instance, an image from an oscilloscope instead of an array containing the values the signal takes for each instant of time.

Since you have a time series problem, why don't you try a RNN based approach?

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.