# 6.3 Prototypes and Criticisms ### MMD-critic 如何做為解釋型 AI ? MMD-critic 提供三面向的解釋- 1. better understand the data distribution 1. building an interpretable model 1. making a black box model interpretable ### Prototype A prototype is a data instance that is representative of all the data. 用一個特定的data instance代表母體 ### Criticism A criticism is a data instance that is not well represented by the set of prototypes. 與Prototypes不像的data instance ##選擇Prototype與Criticism方法: 1. K-medoids 延伸:https://www.itread01.com/content/1546282480.html (K-means 和 K-medoids 之間的差異就類似於一個數據樣本的均值和中位數之間的差異) 2. MMD-critic <=章節主要內容 Maximum Mean Discrepancy (MMD) MMD-critic compares the distribution of the data and the distribution of the selected prototypes. Selects prototypes that minimize the discrepancy between the two distributions. 選擇一組Prototypes,讓此組之Distribution與母體Distribution差異最小 * Theory * 先選你想要的Prototypes與Criticisms數量 * 用貪婪的方式(Greedy Search)找出Prototypes * 用Greedy Search找出Criticisms * Goal : Minimize MMD2 $MMD^2=\frac{1}{m^2}\sum_{i,j=1}^m{}k(z_i,z_j)-\frac{2}{mn}\sum_{i,j=1}^{m,n}k(z_i,x_j)+\frac{1}{n^2}\sum_{i,j=1}^n{}k(x_i,x_j)$ A choice for the kernel(k) is the radial basis function kernel: $k(x,x^\prime)=exp\left(-\gamma||x-x^\prime||^2\right)$ 延伸:Kernel Function https://towardsdatascience.com/kernel-function-6f1d2be6091 * Finding Criticism: $witness(x)=\frac{1}{n}\sum_{i=1}^nk(x,x_i)-\frac{1}{m}\sum_{j=1}^mk(x,z_j)$ ### witness function <公式概念> 用來評估 原始群集與 Prototype 群集之間的相似度 witness = 點X 與原始值的相似度-點X 與經挑選prototype的相似度 <其值代表的意義> 1. 越接近零,表示該點所在處 Prototype 與 原始值分布相似 2. +/-越大方向的值,代表 Prototypes 錯估的方向 這些極端值之絕對值越大,代表點X所在的位置附近的prototype無法被 prototype 好好解釋,適合拿來當criticism * 負越大代表Prototypes不具代表性 * 正越大代表應該納入更多附近的點當Prototypes the witness function gives you the means of evaluating in which empirical distribution the point x fits better. >(語句還原) >> (A) and (B) fits 的程度評估 > the witness function gives you > (A)the means of evaluating empirical > fits better > (B)The distribution which the point x fits in > witness function 與 prototype 挑選方式是相互獨立的 (不管prototype手挑還是演算法挑,witness function都可以評估) ### good sanity check MMD-critic 運用Prototype 與 critics 的分析,可以更了解trainig 資料的分布,也可以了解模型因資料所產生的弱點 Prototype 與 critics 是兩群長的比較不一樣的資料,比較模型對於者兩群的差異,就能知道模型到底夠不夠好 ### 缺點 各種參數的甜蜜點難抓 還沒有好的套件可用 延伸(論文) : https://papers.nips.cc/paper/6300-examples-are-not-enough-learn-to-criticize-criticism-for-interpretability.pdf ###### tags: `重點摘要`