Modelling and analysis of the 8 filters from the "master key filters hypothesis" for depthwise-separable deep networks in relation to idealized receptive fields based on scale-space theory
The research analyzed eight 'master key filters' from ConvNeXt networks using clustering to find essential patterns. It then modeled these filters as idealized receptive fields based on scale-space theory, showing good qualitative similarity. This suggests that learned filters in deep networks can be well approximated by simple discrete scale-space models.