Implement nn_kernel function and visualize samples

Added Python code for neural network kernel and Gaussian process prior samples visualization.

Author: whiteode

chapter_gaussian-processes/gp-priors.md

@@ -148,6 +148,34 @@ In some cases, we can essentially evaluate this covariance function in closed fo
 
 The RBF kernel is _stationary_, meaning that it is _translation invariant_, and therefore can be written as a function of $\tau = x-x'$. Intuitively, stationarity means that the high-level properties of the function, such as rate of variation, do not change as we move in input space. The neural network kernel, however, is _non-stationary_. Below, we show sample functions from a Gaussian process with this kernel. We can see that the function looks qualitatively different near the origin.
 
+```{.python .input}
+def nn_kernel(x1, x2):
+    x1 = x1.flatten()
+    x2 = x2.flatten()
+    N, M = len(x1), len(x2)
+    cov_matrix = np.zeros((N, M))
+    for i in range(N):
+        for j in range(M):
+            tilde_x_i = np.array([1, x1[i]])
+            tilde_x_j = np.array([1, x2[j]])
+            numerator = 2 * np.dot(tilde_x_i, tilde_x_j)
+            term_i = 1 + 2 * np.dot(tilde_x_i, tilde_x_i)
+            term_j = 1 + 2 * np.dot(tilde_x_j, tilde_x_j)
+            arg = np.clip(numerator / np.sqrt(term_i * term_j), -1.0, 1.0)
+            cov_matrix[i, j] = (2 / np.pi) * np.arcsin(arg)
+    return cov_matrix
+
+x_points = np.linspace(-5, 5, 100)
+meanvec = np.zeros(len(x_points))
+covmat = nn_kernel(x_points, x_points)
+prior_samples = np.random.multivariate_normal(meanvec, covmat, size=5)
+
+d2l.plt.plot(x_points, prior_samples.T, alpha=0.7)
+d2l.plt.show()
+
+```
+
+
 ## Summary