[logseq-plugin-git:commit] 2024-08-08T01:54:05.030Z

Author: dino920135

journals/2024_08_07.md

@@ -11,8 +11,13 @@
 		- **in navigation -> not enough -> transfer learning** (predictable behavior)
 		  might not suitable for INS
 	- avoid overfitting by **early stop**
-	- avoid overfitting by use of **L1 or L2 regularization**
-	- for ConvNet
-		- dropout
-- **Overfitting is some how creating stuff**
+	- avoid overfitting by use of **L1 or L2 regularization penalty**
+	- for ConvNet (Or large network)
+		- **dropout**
+		  for forward prop. randomly close some of the neurals (test if the neural is important)
+		- **Batch normalization** (image processing)
+		  Normalization of the result from batch:
+		  Loss of imformation? remove outlier?
+		  ![Normalization in Deep Learning – calculated | content](https://calculatedcontent.com/wp-content/uploads/2017/06/batchnorm2-e1497643748774.png){:height 269, :width 409}
+- **Overfitting is some how creating stuff** (Nerf)
 -