KAG with learnable grids is added (experimental)

README.md

@@ -19,7 +19,8 @@ In addition, I implemented Chebyshev polynomials in KAN.
 - KAC-Net: Utilizing Chebyshev Polynomials in Kolmogorov Arnold Chebyshev Networks
 
 I implemented the Gaussian Radial Basis Functions introduced in [fastkan](https://github.com/ZiyaoLi/fast-kan): 
-- KAG-Net: Utilizing Gaussian radial basis functions in Kolmogorov Arnold Gaussian Networks
+- KAG-Net: Utilizing Gaussian radial basis functions in Kolmogorov Arnold Gaussian Networks (non-trainable grids)
+- KAGL-Net: (Experimental) Utilizing Gaussian radial basis functions in Kolmogorov Arnold Gaussian Networks (trainable grids)
 
 # install
 ```
@@ -52,6 +53,8 @@ model = Chain(KAGnet(2, 10, num_grids=4), KAGnet(10, 1, num_grids=4))
 ```
 In the KAGnet, the grid points are fixed. 
 
+I implemented the Gaussian function with learnable grid points. But this is experimental. You can use ```KAGLnet```. 
+
 # MNIST
 
 ```julia

src/examples.jl

@@ -40,6 +40,11 @@ function MNIST_KAN(; batch_size=256, epochs=20, nhidden=64, polynomial_order=3,
             KAGnet(input_size, nn; num_grids=polynomial_order + 1),
             KAGnet(nn, nclasses; num_grids=polynomial_order + 1)
         )
+    elseif method == "GaussianLearnable"
+        model = Chain(
+            KAGLnet(input_size, nn; num_grids=polynomial_order + 1),
+            KAGLnet(nn, nclasses; num_grids=polynomial_order + 1)
+        )
     else
         error("method = $medhod is not supported")
     end