There are two questions about machine learning that almost every engineer asks within their first year of doing it seriously, and both questions get answered badly by most introductions. The first is: will the model work on data it has never seen? — the question of generalization. The second is: can what the model learned in one setting help it work in a different setting? — the question of transfer. The first is taught with a paragraph about overfitting and a hand-wave about the training/test split; the second is taught with an aside about ImageNet fine-tuning. Neither does justice to the underlying math, and neither makes clear that the two questions are actually the same question — separated by exactly one term in exactly one inequality.
The unifying inequality is the headline of this field guide. For generalization: true error on new data is at most your measured error plus a complexity penalty divided by √M. For transfer: error in the new world is at most your error in the old world plus a divergence term plus a hard-to-control λ. The structure is identical. The only difference is the middle term — the complexity / √M for generalization, the divergence(source, target) for transfer. When the source and target are the same world, the divergence is zero, the transfer inequality collapses, and you are left with generalization. When they are different, the divergence wakes up and dominates everything.
The field guide is built around that structural identity. The landing page is an interactive figure that lets you slide the two worlds apart and watch the divergence term grow. The two deep dives — one for each regime — develop the math, the geometry, and the practical decisions that follow from it. The format is deliberately three pages: the landing page is meant to be read first (it sets up the unifying picture), and the two deep dives can then be read in either order.
This is the natural companion to the Optimization Methods Reference (the methods used to actually minimize the right side of these inequalities), the PyTorch & TensorFlow monograph (where the inequalities show up in practical training-loop discussions), and the RL in Agentic AI post (where transfer specifically is doing a lot of the heavy lifting under the sim-to-real banner).
What it covers
A three-page field guide. About twenty-four minutes of total reading.
The landing page
One inequality, two regimes. The interactive figure with the source and target worlds, the slider that pulls them apart, the live readout of generalization vs. transfer error. The unifying picture that motivates the whole field guide.
Deep dive 01 — Knowledge Generalization
Same task, new examples. The classical generalization inequality (R(h) ≤ R̂(h) + complexity/√M). What “M independent examples” really means (much smaller than your row count, usually). The tug of war between model complexity and evidence. How the gap is actually controlled in practice — regularization, early stopping, model selection. The dozen ways the inequality is misused.
Deep dive 02 — Knowledge Transfer
A different world, borrowed learning. The transfer inequality (ε_target ≤ ε_source + ½·divergence(source, target) + λ). Four things you can carry over from a source task to a target task (representations, hyperparameters, data weights, the model itself). Why transfer succeeds or fails — almost always because of the divergence term. The cases where transfer is the right answer and the cases where training fresh is.
Read it
The landing page is the entry point — it has the interactive figure that makes the unifying picture click. The two deep dives are reachable from the landing page (and cross-link to each other), or you can jump directly to either.
Or jump straight to a deep dive:
↓ Knowledge Generalization — same task, new examples (Part 01 of 02)
↓ Knowledge Transfer — different world, borrowed learning (Part 02 of 02)
All three pages live in the warm-paper editorial layout — rust accent for the unifying inequality, teal for the supporting math, ochre for the practical-notes sections. Each deep dive opens with the inequality it is about, displayed prominently, and develops the intuition and consequences from there.
← Back to Autonomy