Skip to main content

Skills application solution

Instruction and application
In Progress

Skills application solution​

Compare your skills application output to the solution example below provided by Multiverse subject matter experts.

Solution In this activity, you explored optimisation and regularisation techniques to improve a neural network’s generalisation performance on the MedPredict dataset.

  • Optimiser comparison: The Adam optimizer converged faster than SGD and achieved higher validation accuracy in fewer epochs. However, Adam showed slightly higher variance in validation loss, suggesting potential overfitting.

  • Regularization effects: Applying L2 regularisation (λ=10−4\lambda = 10^{-4}λ=10−4) anddropout (rate = 0.2) reduced overfitting, resulting in more stable training and improved validation accuracy. The validation loss curve flattened earlier, indicating better generalisation.

  • Final recommendation: Retain Adam as the optimizer but addearly stopping with a patience of 3 epochs and slightly increasedropout (0.3) to further stabilise results and prevent overfitting.

ModelValidation accuracyValidation lossSGD (no regularisation)0.850.41Adam (no regularisation)0.890.35Adam + L2 + Dropout0.910.30

What this example does well​

  • Clearly compares model performance across optimisers and regularisation settings using both metrics and visual analysis.
  • Connects observed results to training dynamics (speed, stability, generalisation).
  • Provides a justified, evidence-based recommendation that balances model performance and overfitting control.

Tips for applying this skill in your role​

  • Always start with a baseline optimizer and analyse its learning curve before experimenting with adaptive methods.
  • Combine multiple regularisation techniques (e.g., weight decay, dropout, early stopping) for robust generalisation.
  • Evaluate not just accuracy but also training and validation trends — stability is as important as raw performance.

Action item: Reflection​

  • Compare your output to the solution example provided. What did you do well? Where could you improve?
  • How might you adapt your optimisation and regularisation strategy for a larger, noisier dataset in your work context?