Gao is a first year Master student at BUPT, major in Management Science and Engineering, CNN Visualizaiton.

Talks

Attribution Method

There are many methods to explain the predictions of DNNs,and I focus on the problem of assigning an attribution value, sometimes also called “relevance” or “contribution”, to each input feature of a network.

More formally, consider a DNN that takes an input x = [$x_1$, …, $x_N$] ∈ $R^N$ and produces an output S(x) = [$S_1(x)$, …, $S_C(x)$], where C is the total number of output neurons.

Given a specific target neuron c, the goal of an attribution method is to determine the contribution $R_c$ = [$R_1^c$, …, $R_N^c$] ∈ $R^N$ of each input feature $x_i$ to the output $S_c$.

For a classification task, the target neuron of interest is usually the output neuron associated with the correct class for a given sample. When the attributions of all input features are arranged together to have the same shape of the input sample which are called attribution maps.

The attribution maps are usually displayed as heatmaps where red color indicates features that contribute positively to the activation of the target output, and blue color indicates features that have a suppressing effect on it.

Contact

Posts

  1. 第三季5场 AI ML Club 活动纪要 September 08, 2019
Back to Top ↑

Updated: