Minor updates to tutorial_odegen and ml_classical_shadows (#1429)

**Title:**

**Summary:**
- The odegen demo is very sensitive to the initial guess. Updating the
random seed so that the plot looks nicer.
- The classical shadows demo is not compatible with latest jax, and is
using an old interface of sklearn

**Relevant references:**

**Possible Drawbacks:**

**Related GitHub Issues:**

Author: astralcai

demonstrations/ml_classical_shadows.metadata.json

@@ -6,7 +6,7 @@
         }
     ],
     "dateOfPublication": "2022-05-02T00:00:00+00:00",
-    "dateOfLastModification": "2025-01-10T00:00:00+00:00",
+    "dateOfLastModification": "2025-07-10T00:00:00+00:00",
     "categories": [
         "Quantum Machine Learning"
     ],

demonstrations/ml_classical_shadows.py

@@ -27,8 +27,10 @@
 properties from the learned classical shadows. So let's get started!
 
 .. note::
-    This demo is compatible with the latest version of PennyLane and ``neural-tangents==0.6.5``.
+    This demo is compatible with the latest version of PennyLane and ``neural-tangents``.
     The latter is required for building the kernel for the infinite network used in training.
+    As of July 10th, 2025, the latest version of ``neural-tangents`` (v0.6.5) is only compatible
+    with ``jax<0.6.0``.
 
 
 Building the 2D Heisenberg model
@@ -648,7 +650,7 @@ def build_dataset(num_points, Nr, Nc, T=500):
 # from the ``sklearn`` library.
 #
 
-from sklearn.metrics import mean_squared_error
+from sklearn.metrics import root_mean_squared_error
 
 def fit_predict_data(cij, kernel, opt="linear"):
 
@@ -685,8 +687,8 @@ def fit_predict_data(cij, kernel, opt="linear"):
                 best_model = model(hyperparam).fit(X_train, y_train)
                 best_pred = best_model.predict(X_test)
                 best_cv_score = cv_score
-                best_test_score = mean_squared_error(
-                    best_model.predict(X_test).ravel(), y_test_clean.ravel(), squared=False
+                best_test_score = root_mean_squared_error(
+                    best_model.predict(X_test).ravel(), y_test_clean.ravel()
                 )
 
     return (

demonstrations/tutorial_odegen.metadata.json

@@ -6,7 +6,7 @@
         }
     ],
     "dateOfPublication": "2023-12-12T00:00:00+00:00",
-    "dateOfLastModification": "2025-01-28T00:00:00+00:00",
+    "dateOfLastModification": "2025-07-08T00:00:00+00:00",
     "categories": [
         "Optimization",
         "Quantum Computing",

demonstrations/tutorial_odegen.py

@@ -296,7 +296,7 @@ def partial_step(grad_circuit, opt_state, theta):
 
     return thetas, energy
 
-key = jax.random.PRNGKey(0)
+key = jax.random.PRNGKey(888)
 theta0 = jax.random.normal(key, shape=(n_param_batch, tbins * 2))
 
 thetaf_odegen, energy_odegen = run_opt(value_and_grad_jax, theta0)
@@ -313,8 +313,9 @@ def partial_step(grad_circuit, opt_state, theta):
 
 ##############################################################################
 # We see that with analytic gradients (ODEgen), we can reach the ground state energy within 100 epochs, whereas with SPS gradients we cannot find the path
-# towards the minimum due to the stochasticity of the gradient estimates. Note that both optimizations start from the same (random) initial point.
-# This picture solidifies when repeating this procedure for multiple runs from different random initializations, as was demonstrated in [#Kottmann]_.
+# towards the minimum due to the stochasticity of the gradient estimates. Note that the convergence of the optimization is sensitive to the initial guess.
+# In this demonstration, both optimizations start from the same (random) initial point. This picture solidifies when repeating this procedure for multiple
+# runs from different random initializations, as was demonstrated in [#Kottmann]_.
 #
 # We also want to make sure that this is a fair comparison in terms of quantum resources. In the case of ODEgen, we maximally have :math:`\mathcal{R}_\text{ODEgen} = 2 (4^n - 1) = 30` expectation values.
 # For SPS we have :math:`2 N_g N_s = 32` (due to :math:`N_g = 2` and :math:`N_s=8` time samples per gradient that we chose in ``num_split_times`` above). Thus, overall, we require fewer