Hey everybody,
I am a beginner in Neural Networks and I need help with a model for my problem.
I want to classify waveforms into two categories (Binary basically) and I am having a difficult time training an NN for this task.
I have 600 one-dimensional waveforms, each one is 10,000 points long. I have processed them with an RFFT. Each waveform either belongs to class True or False.
Here is a plot of those classes (after RFFT):
As you can see, the waveforms are visibly different. I want to train a Neural Network to classify a given waveform to either class.
I have tried a two-layer CNN but the most I could get was 67% accuracy. I also have tried an LSTM but I think I made lots of mistake in their design.
Here I am, looking for help for a model.
Thank you guys
Important question: what do you mean by 67% accuracy? That is, are you talking about the training accuracy or the accuracy in a separate validation/test set?
If it’s the training accuracy, then there’s either some problem with your implementation (bad learning rate or even a bug in the neural net definition) or your model is too small. In both cases, this does not mean that a CNN is not capable of solving your problem.
If you’re having a high training accuracy (> 90%) but your validation/test accuracy is only 67%, then you are overfitting your data. This could happen if your training set is too small or if there’s no relevant structure in your data that the CNN can capture to classify the waveforms.
I meant the validation/test set. Changing the learning net didn’t make a difference. Increasing the model size made it worse.
However, what finally made it work was some additional topic-specific preprocessing. I now have 87% accuracy.
Thank you for your response. Very very appreciated
