False convolution layer in LTS Shapelet model

[RadhikaaM]

Dear @rtavenar,

First of all thanks for providing ready-to-use package for learning Shapelets in tslearn, that's implemented based on the approach proposed in "Learning Time-Series Shapelets".

Following up the gradient-descent based approach from the paper, I am trying to understand the architecture (layers) used in LTS-Shapelets of tslearn.

I found the first layer which is false convolution layer to be set as trainable=False. Please help me in understanding the significance of this layer. How will it differ the learning, if I remove this layer and connect input directly to LSD layer.

Your insight would be very helpful

Thanks