I am working with the
Pamap2 sensor data for human activity recognition, using
convolutional neural networks having these attributes:
['x-acc-hand','y-acc-hand','z-acc-hand','x-gyr-hand','y-gyr-hand', 'z-gyr-hand','x-mag-hand','y-mag-hand','z-mag-hand','temp-chest', 'x-acc-chest','y-acc-chest','z-acc-chest','x-gyr-chest','y-gyr-chest', 'z-gyr-chest','x-mag-chest','y-mag-chest','z-mag-chest','temp-ankle', 'x-acc-ankle','y-acc-ankle','z-acc-ankle','x-gyr-ankle','y-gyr-ankle', 'z-gyr-ankle', 'x-mag-ankle', 'y-mag-ankle', 'z-mag-ankle']
It is known that neural networks or deep learning has the capability to extract features from raw data for model creation. I however feel it would be important to feed the network with minimal (most relevant raw data) in order not to confuse the network that much.
In traditional machine learning scenario such as
random forest, the algorithm provides the
feature_importances_ model attribute, so one can see (and maybe choose) to create a final model with those important features only.
I am wondering if there's some way to asses the most important attributes of the raw data to feed the network with. That way, if available will results in a robust model than simply feeding my network with these bunch of atttibutes.
May I know if there exist such a workaround in deep learning architectures?