This addresses task in #1.

---

Description of the problem
When training with TRUST_REGION, the $\chi^2$ drops down to less than $1$, resulting in strong over-learning. However, this indicates that both minimisation algorithm and implementation of the $\chi^2$ are working fine. On the other hand, GD vanilla reduces the $\chi^2$ down to $\sim 40$ even after $10000$ iterations. It is required to understand whether this is due to the very small uncertainties of the generated points or there is a bug in the implementation of the $\chi^2$. The latter is less feasible, because the AnalyticChiSquare works when using TRUST_REGION. Still, it is worth checking before drawing any conclusion.

List of things to do and check:

---

CLI:

• train should require only the folder in which data are stored.
• Hence datagen should copy meta.yaml.

---

Consistency check

• #3
• #2
• #8
• #9

---

Random generator:

• Check how the seed in the runcard enters the data generation and the initialisation of replicas
• Check consistency for random numbers in datagen and train

---

NTK scaling:

• Get rid of the NTK_scaling flag. Adopt diagonal learning tensor, take the squared root, and multiply it by the derivative of the network
• Implementing numerical differentiation using functors.

---

Analysis

• Store outputs at each time step to how the network adapts during training
• Modify jupyter notebook for the analysis

---

Code development

• #5
• Implement a configuration manager either from scratch or using [config-cpp](https://github.com/CJLove/config-cpp/tree/master
• Implement testing using Google Test
• Make CMakeList.txt look more professional...
• Introduce the possibility to choose between the LeCun initialisation and the one followed by Jacot et Al..