# 0106 SCCNet with ICA loss ###### tags: `study` ## Setting and Loss * setting為之前參數討論最佳那組 * Lr : 5*1e-4 * batch size : 128 * epoch : 500 * validation ratio : 0.125 * weight decay : 0.1 * $L=\alpha \cdot CE + \beta \cdot (Neg+Rec)$ * $CE$ = Cross Entropy * $Neg$ = Negentropy * $Rec$ = Reconstruction Error * $\alpha \in [0.1, 1], \beta = 1 - \alpha$ * Nonliear in ICA: * $logcosh(S(x))+||W^TS(Wx_i)-x_i||^2_2$ * $S(·)=tanh(·)$ * Scheme : Individual # Analysis ## 1. Mean accuracy of 10 repeats * 只有少數subject會有提升 ![](https://i.imgur.com/bQdTRc0.png) ![](https://i.imgur.com/WpOc59G.png) ![](https://i.imgur.com/2lXDbT1.png) ![](https://i.imgur.com/ftX47Gt.png) ![](https://i.imgur.com/KU7xlwK.png) ![](https://i.imgur.com/lTiz9gJ.png) ![](https://i.imgur.com/ODkRmfV.png) ![](https://i.imgur.com/6sLJgtu.png) ![](https://i.imgur.com/LDXJx1c.png) ## 0~0.1 ### Conclusion * No statistical significance * No Consistency * $\alpha=0.94$ $\beta=0.06$後ICA的loss才會被optimize ### S01 ![](https://i.imgur.com/dnOYS1g.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.55664 | 0.90528 | 0.84570 | 0.32226 | 0.49218 | #### Nonlinear in ICA loss ![](https://i.imgur.com/ik0fKgP.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.44009 | 0.72181 | 0.84570 | 0.95272 | 0.84570 | ### S02 ![](https://i.imgur.com/D2y6wy6.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.69531 | 0.76710 | 0.26203 | 0.08398 | 0.08398 | #### Nonlinear in ICA loss ![](https://i.imgur.com/8ejouo8.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.76953 | 0.16016 | 0.27539 | 0.31394 | 0.19336 | ### S03 ![](https://i.imgur.com/lmKdOTG.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.76690 | 0.84570 | 0.76953 | 0.76953 | 0.40653 | #### Nonlinear in ICA loss ![](https://i.imgur.com/EbpunpM.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.81238 | 0.81238 | 0.37500 | 0.27539 | 0.51356 | ### S04 ![](https://i.imgur.com/JIhjc4f.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.43164 | 0.81189 | 0.72181 | 0.49219 | 0.55327 | #### Nonlinear in ICA loss ![](https://i.imgur.com/stfOPLy.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.37426 | 0.23242 | 0.43164 | 0.67840 | 0.43925 | ### S05 ![](https://i.imgur.com/tx2uTet.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.51430 | 0.27539 | 0.27539 | 0.13864 | 0.55664 | #### Nonlinear in ICA loss ![](https://i.imgur.com/lzlJ6pi.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.19336 | 0.84570 | 0.10974 | 0.62500 | 0.62500 | ### S06 ![](https://i.imgur.com/jvxKCN2.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.08398 | 0.43164 | 0.19336 | 0.32227 | 0.10850 | #### Nonlinear in ICA loss ![](https://i.imgur.com/hLFqqfD.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.01953 | 0.23242 | 0.43164 | 0.27539 | 0.49219 | ### S07 ![](https://i.imgur.com/Q46pGkB.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 1.00000 | 0.55664 | 0.06445 | 0.43164 | 0.13086 | #### Nonlinear in ICA loss ![](https://i.imgur.com/bQRc0bn.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.32227 | 0.13086 | 0.17232 | 0.23242 | 0.44087 | ### S08 ![](https://i.imgur.com/meOMAHg.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.83262 | 0.10547 | 0.39718 | 0.49219 | 0.81254 | #### Nonlinear in ICA loss ![](https://i.imgur.com/YYpvwDM.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.16016 | 0.59363 | 0.92188 | 0.34240 | 0.76953 | ### S09 ![](https://i.imgur.com/zAHuh6F.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.49219 | 0.34198 | 0.37500 | 0.06445 | 0.27539 | #### Nonlinear in ICA loss ![](https://i.imgur.com/DpYMx7H.png) * Signed Rank Test with Basline : p-value | comb 1 | comb 2 | comb 3 | comb 4 | comb 5 | | ------- | ------- |:------- | ------- |:------- | | 0.76953 | 0.85883 | 0.95277 | 0.72564 | 0.34283 |