about max

[yang1332133]

Hi, firstly thanks for your great work! I want to ask how max is manifested in node updates？

[ShunChengWu]

The idea of using FAN is to distribute part feature information in the latent space. The simplest case is using height as the metric to distribute the feature embedding. Let's say a feature has a dimension of three. The description of parts under different heights is stored separately into those three dimensions, e.g.: if height>1: store at feature[0], if height>=-1 and height <=1: store at feature[1], if height<-1: store at feature[2]. The use of max is to select the most significant feature among all feature dimensions, e.g., selecting the most representative part under different heights.

Of course, when projecting everything into high dimensional space and with the implicit learning using NN, this behavior is not guaranteed, especially without any extra constraints. But the above example is the intuition of using FAN and the use of max aggregation. Feel free to try out different aggregation methods.

[yang1332133]

The idea of using FAN is to distribute part feature information in the latent space. The simplest case is using height as the metric to distribute the feature embedding. Let's say a feature has a dimension of three. The description of parts under different heights is stored separately into those three dimensions, e.g.: if height>1: store at feature[0], if height>=-1 and height <=1: store at feature[1], if height<-1: store at feature[2]. The use of max is to select the most significant feature among all feature dimensions, e.g., selecting the most representative part under different heights.

Of course, when projecting everything into high dimensional space and with the implicit learning using NN, this behavior is not guaranteed, especially without any extra constraints. But the above example is the intuition of using FAN and the use of max aggregation. Feel free to try out different aggregation methods.

Thank you for your reply！