SRAI Book 3 · Chapter 1 · PU-B03-C01

Neural Network Foundations

Build neural networks from the artificial neuron upward, then evaluate what their outputs can support through reproducible validation, calibration, subgroup analysis and robustness checks.

Watch the lesson Open the learning resources

01 / LEARNING OUTCOMES

Connect neural-network computation to responsible evidence.

Derive and interpret artificial-neuron and layer computations.
Explain why nonlinear activations create representational capacity.
Implement shape-safe forward propagation and binary cross-entropy.
Separate training, validation and test evidence.
Compare a multilayer perceptron with a credible linear baseline.
Evaluate calibration, thresholds, subgroups, robustness and reproducibility.

02 / EVIDENCE CHAIN

Represent, validate, interpret and govern.

Represent

Transform standardized inputs through weighted layers and nonlinear activations.

Validate

Use protected partitions, a credible baseline and an untouched test set.

Interpret

Examine calibration, thresholds, subgroup evidence and model behavior.

Govern

Document limitations, monitoring, accountability and acceptable use.

03 / CONTROLLED RESULTS

A slightly higher AUC is not a deployment decision.

The canonical notebook compares a logistic baseline with a compact 5-10-1 multilayer perceptron using reproducible synthetic data.

The MLP marginally improves test AUC, while calibration, threshold consequences, subgroup burdens and distribution shift reveal why evaluation must remain multidimensional.

VERIFIED NOTEBOOK RESULTS

Records	1,200
Train / validation / test	780 / 210 / 210
Baseline test AUC	0.7954
MLP test AUC	0.7957
Gradient relative error	2.44e-11
Rerun prediction difference	0.0

REPRESENTATION

A neural network can learn nonlinear scoring functions.

Architecture creates capacity, but it does not determine whether a claim is valid or an operational action is justified.

EVIDENCE

Trust depends on the full chain.

Baselines, calibration, subgroup checks, robustness, reproducibility and governance determine what the results can responsibly support.

04 / SRAI DECISION STANDARD

A neural network earns trust through its evidence chain.

Architecture and tensor shapes are explicit and testable.
Validation selects the checkpoint without exposing the test set.
Performance is compared with a credible baseline.
Threshold consequences and subgroup burdens remain visible.
Distribution shift and independent reruns test robustness.
Synthetic educational evidence is not treated as deployment authorization.

VIDEO LESSON

Watch the complete Lesson 1 presentation.

Open on YouTube ↗

CONTROLLED RESOURCES

Read, reproduce, practise and review.

READ

View the complete GitHub production unit ↗