Foundation Models, Transfer, and Fine-Tuning as a Conceptual Orientation

When labelled data are scarce but unlabelled history is abundant, training a large network from scratch on the label set alone is often the wrong default. Transfer learning reuses parameters or representations learned on a source task (or unlabelled corpus) and adapts them to a target task — the pattern behind modern foundation models. This orientation note, grounded in Ng & Ma's treatment, maps the adaptation choices a quant team must document before comparing a fine-tuned model to a linear baseline: what was pretrained, what stayed frozen, and what labelled budget justified each step. It deliberately stops at that conceptual boundary — not fairness, privacy, or LLM systems engineering.