week4_summary.md

Week 4 Summary: From classical microstructure metrics to learned representations

Cross-Book Summary

1. Classical Microstructure Metrics

• Stereology: Quantifying 3D structure from 2D sections.
• Hand-crafted Descriptors: Intuitive but biased (e.g., aspect ratio).
• Information Bottleneck: Scalar metrics discard crucial subtle details.

2. Learned Representations

• Artificial Neuron: Weighted sum plus non-linear activation.
• Activations: ReLU enables deep learning; Sigmoid/Tanh provide non-linearity.
• Universal Approximation: One hidden layer can approximate any continuous function.
• MLP Topology: Stacked layers build abstract representations.

90-Minute Lecture Strategy

Part 1: Classical Approach

• Standard metrics: Grain size, phase fractions.
• Quantitative Metallography.
• Limitations of hand-crafted features.

Part 2: Feature Engineering vs. Learning

• Concept of Representation.
• Incomplete expert features.
• Shift to learned embeddings.

Part 3: Building Blocks of ML

• Mathematical Neuron.
• Weights and Biases.
• Non-linear activations.
• Forward propagation.

Part 4: Multi-Layer Perceptrons

• Stacking layers for abstraction.
• Hidden layers.
• Universal approximators.

Part 5: Outlook

• Moving to images (CNNs).
• Trusting learned vs. classical metrics.

Quarto Website Update (Summary)

Summary for ML-PC Week 4:

• Transitions from classical stereology to learned representations.
• Reviews limits of hand-crafted microstructure metrics.
• Introduces the artificial neuron, weights, and activations.
• Builds the framework for Multi-Layer Perceptrons (MLPs) to automate feature extraction.