Lesson 0001
Sage — this whole lecture is one question you've already spent real time with, just written in unfamiliar notation: can we ever know a thing itself, or only our accumulated impression of it?
In the Yoga Sutras, samskara names something precise: a mental impression left behind by experience, which then conditions how future experience is perceived. You don't meet the world fresh each time — you meet it through the accumulated residue of everything you've met before. That residue isn't the world. It's a compressed, lawful trace of your encounters with it, and it's what actually does the work of anticipating what comes next.
Hold that idea loosely. It's the exact shape of what STOR 323 calls "training" — we're about to watch the same structure show up wearing Greek letters.
D is accumulated experience — a dataset of past encounters. A is a fixed, lawful procedure: the same rule applied to the same experience always leaves the same impression, exactly as samskara-formation isn't arbitrary. And f̂_D — the fitted model — is the impression itself: a compressed trace of D that will now condition how every future input gets perceived and responded to.
When a new input x arrives, the model doesn't re-examine the world from scratch. It perceives x through f̂_D, the same way you meet a new moment through the samskaras already laid down. That's not a flaw in either system — it's the only way perception of the new can be informed by anything at all.
Below are five real house sales — five encounters. Drag the two sliders until the line (the impression) fits the pattern of the encounters as closely as possible. Notice: you're not being asked to remember each house individually. You're being asked to compress them into a single lawful trace that will now speak for all of them when a new house shows up.
Here is the sharper question underneath all of this. Call the true, complete relationship between house size and price — the real one, the one nobody has ever fully seen — f. What you just fit with the sliders is f̂: an estimate, built from five data points, wearing a little hat to mark it as an estimate and not the thing itself.
f̂ is the finger. f is the moon. Every model you will ever build in this course — every fitted line, every softmax probability, every k-nearest-neighbors vote — is a finger, pointed at a moon nobody gets to look at directly. The formal name for this in STOR 323's language: f̂_D = A(D), an estimate produced from finite data, never equal to the true underlying f. This isn't a limitation the field apologizes for. It's the honest starting premise the entire discipline is built on top of.
Training (forming f̂_D) is one act, finished once. Deployment is different: every new x gets read through the finger f̂_D into a score s, and then someone — not the model — must decide what to do with that score. That decision, δ, is a separate act from the knowing. You'll meet this precisely in a later lesson through the Gita's distinction between right action and its fruits — for now, just notice that the course itself insists on keeping "what do I know" (s) and "what do I do" (a) as two different questions, never collapsed into one.
Answer from memory before checking.
Formal side: An Introduction to Statistical Learning, Ch. 2 — the same train/predict framing, without the philosophy layered on top; useful once the intuition has landed, as a check that it's precise, not just poetic. Philosophical side: browse the fuller Yoga Sutras translations linked above, particularly the sutras on samskara and vasana. See RESOURCES.md for the full list.