Hands-On Machine Learning for Algorithmic Trading

Parameter sharing

As stated earlier, the location of salient features may vary due to distortions or shifts. Furthermore, elementary feature detectors are likely to be useful across the entire image.

CNNs encode these assumptions by sharing or tying the weights for the filter in a given depth slice.

As a result, each depth slice specializes in a certain pattern and the number of parameters is further reduced. Weight sharing does not work as well, however, when images are spatially centered, and key patterns are less likely to be uniformly distributed across the input area.